The landscape

Pick a tool.Any tool.

The workflow matters more than the brand. This page exists so you can choose honestly — including the free options, which matter when a foreign card does not always work.

01 If you are starting today

These are the ones worth your first hour. The rest of the table is for when you know what you want.

Claude Code

Anthropic

Anthropic's agentic coding tool — terminal, IDE, desktop app and browser, all one engine.

Free tier
Nothing. Claude Code requires a Pro, Max, Team, Enterprise or Console (API) account. For a one-evening workshop, Console API keys with a workspace spend limit are the cheap way in.
Best at
The deepest agent loop in a real repository: plan mode, checkpoints and rewind, CLAUDE.md memory, hooks, subagents, MCP servers, git worktrees and a headless mode for scripts. CLAUDE.md, settings and MCP servers are shared across every local surface.
Weakness
No free tier at all — the free claude.ai plan does not include Claude Code, so you need Pro, Max, Team, Enterprise or API credits. Long sessions run into a rolling 5-hour limit plus a weekly one, shared with Claude chat.
curl -fsSL https://claude.ai/install.sh | bash • Windows PowerShell: irm https://claude.ai/install.ps1 | iex

From Kazakhstan Kazakhstan is on the Anthropic supported-countries list for both the API and Claude.ai, so no VPN is needed. There is no individual student price. $20 is roughly 10 200 KZT before bank conversion fees; Kazakhstani Visa/Mastercard cards are reported to work once foreign payments are enabled in the bank app.

Docs ↗

Cursor

Anysphere (part of SpaceX since 14 Aug 2026)

A VS Code fork with a built-in multi-model agent, Tab completion, Plan mode, cloud agents and a CLI.

Free tier
Hobby plan: no credit card required, a limited number of Agent requests, Tab completion included.
Best at
The smoothest GUI agent loop there is: Ctrl/Cmd+I opens the agent, Shift+Tab switches to Plan mode, and you accept or reject every diff per file. One editor with every vendor’s models — Claude, GPT, Gemini, Grok, Composer — and the largest community of rules and tutorials.
Weakness
The usage pool burns fast on frontier models — Cursor's own docs say daily agent users typically spend $60-100/mo. Ownership and pricing changed several times in 2026, and the old free year of Pro for students is gone.
https://cursor.com • CLI: curl https://cursor.com/install -fsS | bash (binary: agent)

From Kazakhstan There is no regional plan for Kazakhstan — the cheap Start plan (₹649/mo) is India-only. The former free year of Pro for students is gone; the students page now promises only promotions at campus and online events.

Docs ↗

GitHub Copilot (agent mode)

GitHub / Microsoft

The default AI layer of VS Code: completions, chat, a local agent mode, code review and a CLI.

Free tier
Free plan: 2 000 completions/month plus limited chat and agent use. Students with GitHub Education get the Copilot Student plan: unlimited completions plus 200 AI credits/month, auto model selection only. Teachers and popular open-source maintainers get Copilot Pro free.
Best at
The cheapest paid agent at $10/mo and zero install inside VS Code, with the tightest GitHub issue-to-PR loop. It also reads everyone else’s instruction files — AGENTS.md and CLAUDE.md are on by default in VS Code.
Weakness
Since 1 Jun 2026 agent work is metered in AI credits (1 credit = $0.01), so the 200 student credits are about $2 of agent usage — heavy agent work is no longer free. Model choice was removed on the Free and Student plans on 24 Jun 2026.
VS Code -> sign in with GitHub -> Chat view -> Agent • CLI: npm install -g @github/copilot

From Kazakhstan GitHub Education is global and needs no card, which makes the Copilot Student plan the most realistic $0 route for a student in Kazakhstan. Apply before the workshop — approval is not instant.

Docs ↗

OpenAI Codex CLI

OpenAI

OpenAI's coding agent in your terminal — one ChatGPT account also unlocks the IDE extension, the desktop app and cloud tasks.

Free tier
Included in ChatGPT Free for quick coding tasks (small allowance). Caution: the GitHub README’s sign-in paragraph names only Plus, Pro, Business, Edu and Enterprise while the pricing page says Free and Go include Codex — test a free account before you rely on it.
Best at
The strongest "one subscription, three surfaces" story, and the CLI itself is open source (Apache-2.0, 125.7k stars). Native AGENTS.md, /init to write it, /review to review your own uncommitted diff, codex exec for scripts and codex resume to reopen a session.
Weakness
Limits are published as ranges per model, not as numbers, and they change with the model you pick. The Free and Go allowances are tiny, the Windows sandbox is labelled experimental, and the npm package is @openai/codex, not codex.
npm install -g @openai/codex • Windows PowerShell: irm https://chatgpt.com/codex/install.ps1 | iex

From Kazakhstan Kazakhstan is on the OpenAI supported-countries list. ChatGPT Go at $8/mo is the cheapest paid plan anywhere that includes Codex.

Docs ↗

02 Which one should I use?

You are a student with a university email and a budget of zero

GitHub Copilot (agent mode)Google JulesGoogle Antigravity

GitHub Education gives you the Copilot Student plan with no card at all: unlimited completions plus 200 AI credits a month. Add Jules for free async pull requests (15 tasks/day), and Antigravity if you are 18 or older and want a bigger free agent.

You have never opened a terminal and you need a demo by Friday

LovableBolt.newv0

Rung 1 gives you a working link in one evening: Lovable for a full-stack app with a backend, Bolt if you want to see the real files and terminal in the browser, v0 for the best-looking UI. Then move up a rung — builders hide the code you actually came to learn.

You already know VS Code and want the shortest path to an agent

GitHub Copilot (agent mode)CursorCline

Same keybindings, no new editor to learn. Copilot is the cheapest paid agent at $10/mo and free for students; Cursor has the smoothest agent loop and every vendor’s models; Cline shows you every tool call and its token cost, which is the best way to understand what an agent really does.

You are comfortable with git and the terminal and want to understand how agents really work

Claude CodeOpenAI Codex CLIAntigravity CLI (agy)opencode

Terminal agents in a real git repository, which is where this course lives. Claude Code is the main track; Codex CLI comes with any ChatGPT plan, even the cheap $8 one; agy is $0 if you are 18+; opencode is open source with free beta models, so you can read the loop instead of guessing it.

You already pay for one AI subscription and do not want a second

Claude CodeOpenAI Codex CLICursor

Do not stack subscriptions — one $20 plan is enough. Claude Pro already includes Claude Code; ChatGPT Plus already includes Codex on three surfaces; and if what you want is one GUI editor with every vendor’s models, that is Cursor Pro at $20.

You have a real repo with tests and want work done while you sleep

Google JulesCodex cloudCopilot coding agentDevin Cloud

Rung 4. Start with Jules because it is free, then use the cloud agent that comes with the subscription you already pay for: Codex cloud, the Copilot coding agent, or Devin Cloud on the $20 Pro plan. All of them need a repo that builds in a clean VM, tests that actually run, and a human who reads the diff before merging.

03 Everything, compared

Grouped by what kind of thing it is. Open one to see the free tier, the honest weakness, and how to start it.

Prompt-to-app builders 4

LovableLovableChat your way to a full-stack web app with a built-in cloud backend and GitHub sync.$25/mo (Pro, 200 credits/month; secondary source)

Free tier5 build credits/day, capped at 30 per month, plus 20 Cloud credits and 4 AI credits a month.

Best atThe fastest path from an idea to a shareable MVP if you have never opened a terminal — the backend and the hosting come with it, so a demo link exists the same evening.

WeaknessCredits vanish into debugging loops, and runtime (Cloud/AI) usage is billed separately after launch. App builders also hide the code, which is exactly the part you came to learn.

From KazakhstanA student discount page exists at lovable.dev/students, but whether Kazakhstani universities qualify is not confirmed — check before you count on it.

https://lovable.devDocs ↗
Bolt.newStackBlitzAn in-browser full-stack dev environment driven by chat — real files, real terminal, zero install.$25/mo (Pro, 10M+ tokens/month, one month of rollover)

Free tier300K tokens/day and 1M tokens/month, with Bolt branding on what you publish.

Best atYou watch the real files and a real terminal while the agent works, so it is the app builder that teaches the most. Nothing to install, so it runs on a weak laptop or a school computer.

WeaknessToken burn grows with the size of the project, so a big app eats the daily allowance quickly.

https://bolt.newDocs ↗
v0VercelPrompt to UI and Next.js apps, with one-click deploy to Vercel.$30/mo (Plus, includes $30 credits and $2 daily login credits)

Free tier$5 of credits a month, 7 messages/day, GitHub sync included.

Best atThe best-looking React/Tailwind UI of any generator, with a design mode and GitHub sync — the fastest way to a landing page you are not ashamed of.

WeaknessNext.js and Vercel centric, and the free allowance is small: 7 messages a day. On 21 Sep 2026 the Plus card showed "$30" next to a "$90" figure, so re-check the price before you plan a budget.

From KazakhstanA student plan exists, but eligibility from Kazakhstan is not confirmed.

https://v0.appDocs ↗
Replit AgentReplitA cloud IDE plus an agent that builds, hosts and gives you a database — from any machine, even a phone.$20/mo (Core, $18 annual, includes $20 of model credits)

Free tierStarter: daily Agent credits with a monthly cap, 1 published app (link expires after 30 days), 2 GB, Lite build mode only.

Best atEverything in one place: code, database and deploy. It is the only realistic option when your only device is a Chromebook, a borrowed computer or a phone.

WeaknessEffort-based agent pricing is hard to predict, Full build and Plan mode are paid-only, and on Starter a published app’s link expires after 30 days.

https://replit.comDocs ↗

AI IDEs and editor agents 4

CursorAnysphere (part of SpaceX since 14 Aug 2026)A VS Code fork with a built-in multi-model agent, Tab completion, Plan mode, cloud agents and a CLI.$20/mo (Pro).cursor/rules/*.mdc (also reads AGENTS.md, including nested files)

Free tierHobby plan: no credit card required, a limited number of Agent requests, Tab completion included.

Best atThe smoothest GUI agent loop there is: Ctrl/Cmd+I opens the agent, Shift+Tab switches to Plan mode, and you accept or reject every diff per file. One editor with every vendor’s models — Claude, GPT, Gemini, Grok, Composer — and the largest community of rules and tutorials.

WeaknessThe usage pool burns fast on frontier models — Cursor's own docs say daily agent users typically spend $60-100/mo. Ownership and pricing changed several times in 2026, and the old free year of Pro for students is gone.

From KazakhstanThere is no regional plan for Kazakhstan — the cheap Start plan (₹649/mo) is India-only. The former free year of Pro for students is gone; the students page now promises only promotions at campus and online events.

https://cursor.com • CLI: curl https://cursor.com/install -fsS | bash (binary: agent)Docs ↗
Google AntigravityGoogleGoogle's agent-first dev platform: a VS Code-style IDE plus a multi-agent desktop app.$19.99/mo via Google AI Pro (secondary source).agents/rules/ (also reads AGENTS.md and GEMINI.md)

Free tierIndividual plan at $0: weekly rate limits, unlimited tab completions, and the multi-vendor model list above. Higher limits ride on a Google AI Pro or Ultra subscription; there is no Antigravity-only plan.

Best atThe most generous free agent access available right now. The $0 Individual plan lists Gemini 3.8/3.7/3.6 Flash, Gemini 3.1 Pro, Claude Sonnet and Opus 4.6 and gpt-oss-120b — non-Google models included — with unlimited tab completions.

WeaknessNo published source, interactive Google sign-in, and weekly limits that are never stated as numbers (check with /usage). Google cut its free tiers twice in 2026, so treat "free" as having a shelf life. Unavailable to users under 18.

From KazakhstanKazakhstan is explicitly on the supported list (so are Kyrgyzstan and Uzbekistan), personal Google accounts only. It is unavailable to users under 18, so school pupils need Copilot Student or Cursor Hobby instead.

https://antigravity.googleDocs ↗
Devin Desktop (ex-Windsurf)CognitionThe Windsurf editor, rebranded on 2 Jun 2026: the Devin Local agent, an Agent Command Center, and other agents over the Agent Client Protocol.$20/mo (Pro).windsurf/rules/ or .devin/rules/ (also reads AGENTS.md)

Free tierFree $0: a light quota to code with agents, limited model availability, unlimited Tab completions and inline edits — but no Devin Cloud.

Best atOne $20 plan covers both the local IDE agent and the Devin Cloud agent, so you get rung 2 and rung 4 for one price. Even the free tier keeps unlimited Tab completions and inline edits.

WeaknessThird name in two years (Codeium -> Windsurf -> Devin Desktop), so almost every tutorial you find is stale and windsurf.com docs now redirect to docs.devin.ai. The free quota is light and the model list on it is limited.

From KazakhstanBesides cards, checkout accepts Apple Pay, Google Pay, WeChat Pay and Alipay — useful when a Kazakhstani card is refused by the processor.

https://devin.ai — download for Mac, Windows or Linux; import your VS Code or Cursor settings during onboardingDocs ↗
AWS KiroAmazon Web ServicesA spec-driven agentic IDE: requirements -> design -> tasks, plus a CLI, a web app and an iOS TestFlight build.$20/mo (Pro, 1 000 credits).kiro/steering/*.md (starter files: product.md, tech.md, structure.md)

Free tier50 credits/month with open-weight models and Claude Sonnet 4.5 (the free list also names Qwen3 Coder Next, DeepSeek 3.2 and MiniMax M2.1). Credits do not roll over.

Best atIt forces the discipline this whole course teaches by hand: you cannot skip the spec. Steering files and hooks make the project rules explicit instead of hoping the agent guesses them.

Weakness50 free credits a month is a demo, not a workflow; the sign-in is AWS-flavoured (AWS Builder ID or IAM Identity Center); and the model list is narrower than Cursor’s.

https://kiro.devDocs ↗

Editor extensions 3

GitHub Copilot (agent mode)GitHub / MicrosoftThe default AI layer of VS Code: completions, chat, a local agent mode, code review and a CLI.$10/mo (Pro, includes $15 of AI credits).github/copilot-instructions.md (also reads AGENTS.md and CLAUDE.md in VS Code)

Free tierFree plan: 2 000 completions/month plus limited chat and agent use. Students with GitHub Education get the Copilot Student plan: unlimited completions plus 200 AI credits/month, auto model selection only. Teachers and popular open-source maintainers get Copilot Pro free.

Best atThe cheapest paid agent at $10/mo and zero install inside VS Code, with the tightest GitHub issue-to-PR loop. It also reads everyone else’s instruction files — AGENTS.md and CLAUDE.md are on by default in VS Code.

WeaknessSince 1 Jun 2026 agent work is metered in AI credits (1 credit = $0.01), so the 200 student credits are about $2 of agent usage — heavy agent work is no longer free. Model choice was removed on the Free and Student plans on 24 Jun 2026.

From KazakhstanGitHub Education is global and needs no card, which makes the Copilot Student plan the most realistic $0 route for a student in Kazakhstan. Apply before the workshop — approval is not instant.

VS Code -> sign in with GitHub -> Chat view -> Agent • CLI: npm install -g @github/copilotDocs ↗
ClineCline (open source)An open-source agent inside VS Code, JetBrains, Cursor or Antigravity — plus a CLI, a Kanban board and an SDK.free.clinerules

Free tierThe extension and the CLI are free and open source (Apache-2.0). You pay only for inference: your own key, the Cline provider at cost, or a ClinePass subscription.

Best atTotal transparency: you see every tool call and what each one costs in tokens, which makes it the best tool for actually understanding an agent loop. Plan and Act modes, and any model — including local ones.

WeaknessBring your own key means a card with API billing, and frontier-model API usage costs more than a flat $20 plan for the same amount of work.

npm install -g cline (Node 20+), then: cline auth • or install the extension from the VS Code marketplaceDocs ↗
Roo Code (shut down)Roo Code Inc.Shut down on 15 May 2026 — listed here so a 2025 tutorial does not send you into a dead tool.free

Free tierNot available — the extension is gone and the repository is archived.

Best atNothing today. Its value is historical: it was the most popular Cline fork, and its README now points you to Cline or to the community fork ZooCode.

WeaknessThe extension was shut down and the repository archived on 15 May 2026, so everything you read about it is out of date.

Docs ↗

Terminal agents 6

Claude CodeAnthropicAnthropic's agentic coding tool — terminal, IDE, desktop app and browser, all one engine.$20/mo (Pro; $17/mo billed annually)CLAUDE.md (since v2.1.277 it reads AGENTS.md instead when there is no CLAUDE.md)

Free tierNothing. Claude Code requires a Pro, Max, Team, Enterprise or Console (API) account. For a one-evening workshop, Console API keys with a workspace spend limit are the cheap way in.

Best atThe deepest agent loop in a real repository: plan mode, checkpoints and rewind, CLAUDE.md memory, hooks, subagents, MCP servers, git worktrees and a headless mode for scripts. CLAUDE.md, settings and MCP servers are shared across every local surface.

WeaknessNo free tier at all — the free claude.ai plan does not include Claude Code, so you need Pro, Max, Team, Enterprise or API credits. Long sessions run into a rolling 5-hour limit plus a weekly one, shared with Claude chat.

From KazakhstanKazakhstan is on the Anthropic supported-countries list for both the API and Claude.ai, so no VPN is needed. There is no individual student price. $20 is roughly 10 200 KZT before bank conversion fees; Kazakhstani Visa/Mastercard cards are reported to work once foreign payments are enabled in the bank app.

curl -fsSL https://claude.ai/install.sh | bash • Windows PowerShell: irm https://claude.ai/install.ps1 | iexDocs ↗
OpenAI Codex CLIOpenAIOpenAI's coding agent in your terminal — one ChatGPT account also unlocks the IDE extension, the desktop app and cloud tasks.$8/mo (ChatGPT Go); $20/mo (Plus)AGENTS.md (global ~/.codex/AGENTS.md first, then every directory from repo root down to your cwd; 32 KiB cap)

Free tierIncluded in ChatGPT Free for quick coding tasks (small allowance). Caution: the GitHub README’s sign-in paragraph names only Plus, Pro, Business, Edu and Enterprise while the pricing page says Free and Go include Codex — test a free account before you rely on it.

Best atThe strongest "one subscription, three surfaces" story, and the CLI itself is open source (Apache-2.0, 125.7k stars). Native AGENTS.md, /init to write it, /review to review your own uncommitted diff, codex exec for scripts and codex resume to reopen a session.

WeaknessLimits are published as ranges per model, not as numbers, and they change with the model you pick. The Free and Go allowances are tiny, the Windows sandbox is labelled experimental, and the npm package is @openai/codex, not codex.

From KazakhstanKazakhstan is on the OpenAI supported-countries list. ChatGPT Go at $8/mo is the cheapest paid plan anywhere that includes Codex.

npm install -g @openai/codex • Windows PowerShell: irm https://chatgpt.com/codex/install.ps1 | iexDocs ↗
Antigravity CLI (agy)GoogleThe terminal agent that replaced Gemini CLI for everyone who is not an enterprise customer.$19.99/mo via Google AI Pro (secondary source).agents/rules/ (also reads AGENTS.md and GEMINI.md in the working directory)

Free tierThe same $0 Individual plan as Antigravity; run /usage to see the remaining quota per model.

Best atA $0 terminal agent with real autonomy controls: /goal runs a task to completion without asking at every step, /grill-me makes it interview you before it builds, /usage shows what is left per model, and /browser and /schedule are built in.

WeaknessSame caveats as the IDE: no published source, 18+ only, opaque weekly limits, and an interactive Google sign-in that is awkward on a headless machine (it can print an authorization URL instead).

From KazakhstanSupported in Kazakhstan with a personal Google account, 18+ only. If the venue Wi-Fi blocks the localhost OAuth callback, agy prints an authorization URL you can open on another device.

curl -fsSL https://antigravity.google/cli/install.sh | bash • PowerShell: irm https://antigravity.google/cli/install.ps1 | iex • then: cd project && agyDocs ↗
Gemini CLIGoogleThe open-source Gemini terminal agent — enterprise-only in practice since 18 Jun 2026.freeGEMINI.md (also reads AGENTS.md)

Free tierNone for consumers. Only Gemini Code Assist Standard/Enterprise, Google Cloud or a paid Gemini Enterprise Agent Platform key still work.

Best atStill the reference open-source implementation of a terminal agent (Apache-2.0, ~107k stars), and a perfectly good tool if your university or employer has Gemini Code Assist Standard/Enterprise or a paid key.

WeaknessOn 18 Jun 2026 it stopped serving free users and Google AI Pro/Ultra subscribers; consumers were moved to Antigravity CLI. The README still advertises 1 000 requests/day with a Google login while the docs site contradicts it — do not build a workshop on it.

npm install -g @google/gemini-cliDocs ↗
opencodeAnomalyThe most popular open-source alternative to Claude Code: a TUI, a desktop app and an IDE extension, provider-agnostic.freeAGENTS.md (/init generates it)

Free tierThe tool itself is free, plus a rotating set of free beta models on Zen (today: Big Pickle, MiMo-V2.5, Nemotron 3 variants, Muse Spark 1.3 contributor tier). Zen is otherwise pay-as-you-go and auto-reloads $20 when the balance drops below $5 — you can turn that off.

Best atGenuinely $0 experiments through the rotating free models on Zen, switching models mid-project, and source you can read to learn how an agent actually works. /connect picks a provider, /init writes your AGENTS.md.

WeaknessFree Zen models are weak and unstable and may use your prompts for training; good models mean your own API key. Since 19 Mar 2026 you cannot sign in with a Claude Pro/Max subscription (removed per Anthropic legal requests), and an unauthenticated RCE was disclosed in Jan 2026 — keep it updated.

curl -fsSL https://opencode.ai/install | bash • Windows: choco install opencode or scoop install opencodeDocs ↗
AiderAider (open source)The original open-source terminal pair programmer: git-native, with a repo map and an architect/editor mode.freeAGENTS.md

Free tierFree and open source; you pay only for the model through your own API key.

Best atIt teaches git discipline better than anything else, because it commits every change automatically — every step is reversible. Tiny, scriptable, and happy with local models.

WeaknessDevelopment has stalled: the last tagged release is v0.86.0 from 9 Aug 2025 and its loop is simple next to a 2026 agent. Treat it as legacy — good for learning the idea, not for daily work.

python -m pip install -U aider-chatDocs ↗

Async and cloud agents 4

Codex cloudOpenAIHand a task to Codex in the cloud: it works in its own environment and comes back with a pull request.$20/mo (ChatGPT Plus)AGENTS.md

Free tierShares the Codex allowance of your ChatGPT plan — usable to try on Free, not enough to work with.

Best atSame account, same AGENTS.md and same habits as the CLI, so there is nothing new to learn. Cloud tasks plus GitHub PR review make it the least painful step onto rung 4.

WeaknessIt needs a repo that builds in a clean environment, real tests and a human who reads the diff. It draws on the same Codex allowance as your local work, and on Free or Go that allowance is tiny.

From KazakhstanKazakhstan is on the OpenAI supported-countries list, so no VPN is needed; ChatGPT Go at $8/mo is the cheapest entry that includes Codex.

codex cloud (from the Codex CLI)Docs ↗
Copilot coding agentGitHubAssign a GitHub issue to Copilot; it works in the cloud and returns a pull request.$10/mo (Pro, includes $15 of AI credits)AGENTS.md or .github/copilot-instructions.md

Free tierDraws on the same AI credit balance as the rest of Copilot (1 credit = $0.01); the Student plan gives 200 credits/month.

Best atIt lives where the work already is — issues, pull requests, code review — so your teammates do not have to install anything or learn a new tool. It reads the repository instruction files you already keep.

WeaknessMetered in AI credits like the rest of Copilot, so one long agent run is real money — 200 student credits are about $2. And it only pays off on a repo where CI actually passes.

Assign an issue to Copilot in your GitHub repositoryDocs ↗
Google JulesGoogleConnect a GitHub repo, describe a task, and Jules works in a cloud VM and opens a pull request.$19.99/mo via Google AI Pro (100 tasks/day, 15 concurrent)AGENTS.md

Free tier15 tasks/day, 3 concurrent, Gemini 2.5 Pro, on a rolling 24-hour window.

Best atThe best free introduction to async agents: 15 tasks a day, 3 at a time, no card. Ideal for chores — tests, dependency bumps, small bug fixes — and it reads your AGENTS.md.

WeaknessThe free tier runs an older model (Gemini 2.5 Pro), your repo has to build in a clean VM, and feedback is much slower than a local agent.

From KazakhstanThe free tier needs only a Google account and a GitHub repo — no card at all. Paid tiers are sold through Google AI Pro/Ultra and only for personal @gmail accounts.

https://jules.google — sign in with Google, then connect GitHubDocs ↗
Devin CloudCognitionThe autonomous cloud software engineer: give it a ticket, get a pull request.$20/mo (Pro).devin/rules/ (also reads AGENTS.md)

Free tierNot included in the Free plan — cloud agents start at Pro. The old Core ($20 + $2.25 per ACU) and Team ($500) plans were retired in Apr 2026.

Best atBuilt for delegating a whole ticket rather than a single edit, and it now comes inside the $20 Pro plan together with the desktop agent.

WeaknessIt still needs a repo with tests and a human reviewer to be worth it. The old "$500 per month Devin" reputation lingers, so people dismiss it before they check the current price.

https://devin.aiDocs ↗

04 The models underneath

The environment and the model are two different choices

Most tools let you swap the model. A cheap model in a good harness often beats an expensive model in a bad one — and knowing the difference is what stops you paying for the wrong thing.

Flagship

Claude Fable 5.1

Anthropic

The most capable model Anthropic has released widely. Long agent runs across a big codebase, adaptive thinking always on, and the full 1M-token window billed at the normal per-token rate — a 900k-token request costs the same per token as a 9k one.

  • 1000k context
  • $10 / $50 in / out per 1M tokens

The most expensive option on this page: five times Sonnet 5 for input, five times for output. Anthropic own documentation says to start with Opus 5 for most workloads, so reach for Fable only when Opus visibly fails. Claude Mythos 5.1 is the same capability at the same price but is limited to Project Glasswing participants. Cache reads are 2.5% of input here, cheaper than the usual 10%.

No standing free tier. New API accounts get a small amount of free credits and that is all.

Claude Opus 5

Anthropic

The default recommendation in the Anthropic docs: start here for most workloads. Same 1M window and same 128k max output as Fable 5.1, at half the price.

  • 1000k context
  • $5 / $25 in / out per 1M tokens

There is also a fast mode research preview billed at Fable prices ($10 / $50), so check which mode your client actually sends before you assume the cheap rate. Anthropic commits to keeping Opus 5 available until at least 24 July 2027.

No standing free tier. New API accounts get a small amount of free credits and that is all.

GPT-6 Astra

OpenAI

Described by OpenAI as their most capable model, and tied with Claude Fable 5.1 at the top of the independent Artificial Analysis Intelligence Index (53 at max effort). Knowledge cutoff 30 April 2026.

  • 1050k context
  • $10 / $50 in / out per 1M tokens

Same headline price as Fable 5.1, plus a 2x multiplier on long-context requests — so a genuinely big prompt costs double what the table says. Cached input is 10% of the normal rate, which matters a lot inside an agent loop.

Codex, the OpenAI coding agent, is listed as included in ChatGPT Free for quick coding tasks — but the Codex README only names paid plans, so test a free account before you rely on it in front of a room. On Plus the allowance is roughly 5 to 45 Astra messages per 5 hours.

Gemini 3.1 Pro Preview

Google

The newest Gemini Pro and the strongest Gemini for hard reasoning, with the same 1,048,576-token window as 3.8 Flash. Cheap for a flagship compared with Astra or Fable.

  • 1049k context
  • $2 / $12 in / out per 1M tokens

Two things to know. It is still labelled preview more than half a year after its February 2026 release. And it is the one Gemini with no free tier — above a 200k prompt the price also changes to $4 / $18, so long context is not flat-rate here the way it is at Anthropic.

Not in the API. But Google Antigravity lists Gemini 3.1 Pro on its free Individual plan, and Kazakhstan is on the supported country list — the catch is that Antigravity is not available to anyone under 18.

Muse Spark 1.3

Meta

The current Meta frontier model, released 2 September 2026 by Meta Superintelligence Labs as the replacement for Llama inside Meta products. Very cheap for a flagship, with generous standard rate limits (3,000 requests and 4M tokens per minute).

  • $1.25 / $4.25 in / out per 1M tokens

Closed weights. This is the correction to the most common outdated claim in the room: Meta is no longer the open-source Llama company at the frontier. Zuckerberg has said Muse Spark 1.2 will be released open-weight, but that is announced, not shipped. The ~1M context figure circulating online is not confirmed in the Meta docs, so it is left blank here. See the Contributor tier below for the cheap version and its price in data.

Qwen3.8-Max

Alibaba (Qwen)

The Alibaba flagship: up to 1M context at $2 / $6, which undercuts the Western flagships by a factor of five on input and eight on output. Output being only three times input is unusually flat, and it helps on generation-heavy work.

  • 1000k context
  • $2 / $6 in / out per 1M tokens

Claims that it is a 2.4 trillion parameter mixture of experts are unverified — do not repeat them from a stage. The prices and the free quota here come from the Singapore international deployment of Alibaba Model Studio.

1M free tokens, valid for 90 days, on Alibaba Cloud Model Studio.

Grok 4.7

xAI

The family you meet mainly through Cursor, whose model docs list Grok 4.5, 4.6 and 4.7 including 500k-context and Fast variants next to Claude, GPT and Gemini.

    The price and context cells are deliberately blank: no first-party figure was verified for this research, and a model you cannot price is a model you cannot budget. Worth knowing the context — Cursor has belonged to SpaceX since 14 August 2026, and its pricing page now advertises Grok limits and Grok Bot access alongside the other vendors. Notice when a tool has a reason to steer you towards one engine.

    The Cursor Hobby plan is free with no credit card and a limited number of agent requests. The Cursor pricing page advertises generous limits for Grok without saying on which plan, so check before you count on it.

    Balanced

    Claude Sonnet 5

    Anthropic

    The workhorse of everyday vibecoding: a full 1M context and real agent quality at a fifth of the Fable 5.1 input price. If you are picking one Claude model to live in, this is usually it.

    • 1000k context
    • $2 / $10 in / out per 1M tokens

    The $2 / $10 was a launch promotion that became the permanent price — the announced rise to $3 / $15 on 1 September 2026 was cancelled. Older Sonnet 4.6 and 4.5 are still $3 / $15, so do not copy a price table written in 2025.

    No standing free tier. New API accounts get a small amount of free credits and that is all.

    GPT-5.6 Sol

    OpenAI

    The balanced OpenAI workhorse for agent coding: 1.05M context, a lot cheaper than Astra, and the model most Codex tutorials assume. Knowledge cutoff 16 February 2026.

    • 1050k context
    • $4 / $20 in / out per 1M tokens

    The $4 / $20 is promotional pricing available at least through 21 November 2026, and no regular price is printed anywhere. Blogs quoting $5 / $30 for Sol are misreading the separate GPT-5.5 row — that is a different model with under 272k context. Budget for a rise.

    Reachable through Codex on a paid ChatGPT plan; on Plus the allowance is roughly 10 to 100 Sol messages per 5 hours. The ChatGPT Go plan at $8 is the cheapest paid plan that includes Codex.

    GPT-5.6 Terra

    OpenAI

    Half the input price of Sol with the same 1.05M window. A sensible default the day the Sol promotion ends, and the tier where a Codex subscription stretches furthest.

    • 1050k context
    • $2 / $12 in / out per 1M tokens

    In Codex on a Plus plan, Terra allows roughly 25 to 200 messages per 5 hours against 5 to 45 for Astra — the cheaper model buys you five times more attempts, and attempts are what agent work is made of.

    Gemini 3.8 Flash

    Google

    The Google changelog calls it their most intelligent Flash model, engineered for long-horizon software engineering. GA since 2 September 2026, 1,048,576 tokens in and 65,536 out, and it has a free tier. For a student in Kazakhstan this is the practical default.

    • 1049k context
    • $0.75 / $3.75 in / out per 1M tokens

    The price is promotional through 31 December 2026; from 1 January 2027 it doubles to $1.50 / $7.50. Google has cut or repriced free access twice in 2026, so do not design a product around this tier lasting forever.

    Yes. The Gemini API free tier is the only first-party frontier-lab API that works without a card, which is why it is the right choice for a workshop. Google no longer publishes the rate limits in the docs — check your own numbers in Google AI Studio and do not quote figures from old blog posts.

    Muse Spark 1.3 (Contributor tier)

    Meta

    Exactly the same model as above, at a twelfth of the input price and a twentieth of the output price. On paper the best value in this entire catalogue.

    • $0.1 / $0.2 in / out per 1M tokens

    Read why it is cheap. The Meta page describes it as heavily discounted pricing in exchange for permission to use your prompts and completions to train future Meta models. Rate limits are tighter too: 100 requests per minute against 3,000 on the standard tier. This is the clearest lesson in the catalogue — if it is nearly free, your data is the price. Never point it at a client codebase, a private repo, or anything covered by an NDA.

    The opencode Zen gateway has carried the Muse Spark 1.3 contributor tier in its rotating set of free beta models. The same warning applies twice over: free models there may use your prompts for training.

    Qwen3.7-Plus

    Alibaba (Qwen)

    The Qwen middle tier, up to 1M context, priced below most small models from Western labs.

    • 1000k context
    • $0.4 / $1.6 in / out per 1M tokens

    Input pricing is stepped, not flat: $0.40 per million up to 256k, then $1.20 per million from 256k to 1M. Filling a big window triples the input rate here, which is the opposite of how Anthropic bills long context.

    1M free tokens, valid for 90 days, on Alibaba Cloud Model Studio.

    Small & fast

    Claude Haiku 4.5

    Anthropic

    Cheap and fast for narrow, well-defined jobs: classifying, extracting, summarising, and running as a sub-agent under a bigger model.

    • 200k context
    • $1 / $5 in / out per 1M tokens

    The smallest window in the Anthropic line at 200k, and Anthropic only commits to keeping it alive until 15 October 2026 — do not build a demo you plan to show next year on it. Its knowledge is also the oldest of the four (training data to July 2025).

    No standing free tier. New API accounts get a small amount of free credits and that is all.

    GPT-5.6 Luna

    OpenAI

    Volume work inside the OpenAI family: classification, small edits, sub-agents, bulk rewriting. A fiftieth of the Astra input price with the same 1.05M window.

    • 1050k context
    • $0.2 / $1.2 in / out per 1M tokens

    Older small models are still on the price list (GPT-5.4-mini at $0.75 / $4.50, GPT-5-mini at $0.25 / $2.00) and are simply worse value for the same job. Inside Codex on Plus, Luna carries by far the biggest allowance: roughly 250 to 2,000 messages per 5 hours.

    Gemini 3.5 Flash-Lite

    Google

    The small, sub-agent tier of the current Gemini generation, and it still has a free tier: routing, classification, cheap bulk calls, the boring half of an agent pipeline.

    • $0.3 / $2.5 in / out per 1M tokens

    Google prints context and output limits only on individual model pages, and only Gemini 3.8 Flash and 3.1 Pro Preview were confirmed. Open the model page before you design around a context number for this one.

    Yes, on the Gemini API free tier — no card needed. Limits are shown in Google AI Studio, not in the docs.

    Gemini 2.5 Flash-Lite

    Google

    The cheapest model on this entire page, and it has a free tier — which makes it a fine place to make your very first API call and watch a token counter move.

    • $0.1 / $0.4 in / out per 1M tokens

    An older 2.5-generation model. The Flash models Google describes as built for software engineering are the 3.x ones — use this for simple text work and move up when the task involves a codebase.

    Yes, on the Gemini API free tier — no card needed.

    Codestral

    Mistral

    A dedicated code-completion model — the tab-completion kind of AI rather than the agent kind. Useful to know the category exists, because completion and agent work are priced and judged differently.

    • $0.3 / $0.9 in / out per 1M tokens

    Premier licence, meaning closed weights, unlike most of the Mistral line. Context size is not printed on the pricing page, so it is blank here.

    Qwen3.8-Flash

    Alibaba (Qwen)

    Up to 1M context for $0.15 per million input. As a cheap tier it is unusual in keeping the full context window rather than cutting it.

    • 1000k context
    • $0.15 / $0.47 in / out per 1M tokens

    Qwen tokenizers are among the best documented for Turkic languages — a 2026 study found Qwen3.5 and GPT-5.2 consistently achieved the best Turkic coverage, which matters directly for Kazakh text.

    1M free tokens, valid for 90 days, on Alibaba Cloud Model Studio.

    Open weights

    gpt-oss 20B

    OpenAI

    The open-weight OpenAI model that a real laptop can actually hold: a 14 GB download that Ollama says runs with as little as 16 GB of memory. The cheapest way to feel how an LLM behaves when nothing leaves your machine.

    • open weights

    The big sibling gpt-oss:120b is a 65 GB download and wants an 80 GB GPU, so that one is cloud only. Be honest about what local buys you: privacy, offline work, zero cost and understanding — not frontier agent quality.

    Free: pull it with Ollama and run it offline. Google Antigravity also lists gpt-oss-120b on its free Individual plan, so you can try the big one without hardware.

    Gemma 4

    Google

    The open-weight Google family, and the most laptop-friendly one on this page: sizes on Ollama run from e2b at 7.2 GB through 12b at 7.6 GB to 31b at 20 GB, with vision, tool use and thinking.

    • open weights

    Context is listed as 128k to 256k depending on size. Be careful with blog guides claiming the e2b size needs only about 3 GB of RAM — that conflicts with the 7.2 GB download, so test it on a real laptop before you promise anything to a room full of students.

    Free to download and run via Ollama.

    DeepSeek V4.1-Flash

    DeepSeek

    Serious quality at roughly a tenth of Western flagship prices: 1M context, 384k max output, thinking and non-thinking modes, vision, tool calls and JSON output. It also speaks the Anthropic API format, so Anthropic-compatible clients can be pointed straight at it.

    • 1000k context
    • $0.3 / $1.2 in / out per 1M tokens
    • open weights

    The prices shown are peak. Off-peak is half: peak runs 01:00-04:00 and 06:00-10:00 UTC on weekdays, which in Kazakhstan (UTC+5) means 06:00-09:00 and 11:00-15:00 local — your evenings and weekends are the cheap window. And read the second half of open weights carefully: the model card describes a 552B backbone, so open does not mean it fits on your machine.

    The weights are MIT-licensed on Hugging Face, so it is free if you own hardware for a 552B-parameter model. The hosted API is not free, but it is the cheapest serious API here.

    DeepSeek V4-Pro

    DeepSeek

    The bigger DeepSeek model (V4-Pro-0813, listed at 1.7T parameters on Hugging Face) with the same 1M context and 384k max output, still cheaper than any Western flagship.

    • 1000k context
    • $1.32 / $3.96 in / out per 1M tokens
    • open weights

    No vision, unlike Flash. Same peak and off-peak split. It costs about four times what DeepSeek Flash does, so start with Flash and only move up if you can show that Flash is the thing failing.

    Muse Glimmer 30B

    Meta

    The small open-weight model Meta did ship: 30B parameters under Apache 2.0, released 10 August 2026, with a 128k context. Proof that Meta ships nothing open any more is also wrong.

    • 128k context
    • open weights

    The download on Ollama is 18-19 GB. By the same rule of thumb that lets gpt-oss:20b (14 GB) run on 16 GB of memory, you want noticeably more than 16 GB here. 30B is not a frontier model — use it to learn how these things behave, not to ship a product.

    Free to download and run via Ollama, where it already has over 200 thousand pulls.

    Llama 4 Scout

    Meta

    The last open Llama, released 5 April 2025: 17B active parameters of 109B total, and a 10M-token context window that is still the largest number on this page by a factor of ten.

    • 10000k context
    • open weights

    History, not a recommendation. Meta moved to closed weights in 2026 and there has been no Llama 5. Its sibling Maverick is 17B active of 400B total with 1M context. One more reason it matters to us: Llama-family tokenizers are the worst of the lot for Kazakh, about 4.73 tokens per word against 2.56 for Russian and 2.23 for Turkish.

    Mistral Medium 3.5

    Mistral

    The serious European open-weight option: a dense 128B model under a modified MIT licence, released 28 April 2026. If the requirement is that data never leaves your own servers, this is the tier that can actually do the work.

    • 256k context
    • $1.5 / $7.5 in / out per 1M tokens
    • open weights

    The 256k context is the smallest among the big labs — everyone else here sits at 1M, and that gap is felt immediately in agent work on a large repo. Self-hosting is also not a laptop story: the Ollama download is 80 GB.

    The weights are downloadable under a modified MIT licence, so self-hosting is free of licence cost. The hosted API is paid.

    Mistral Small 4

    Mistral

    Apache 2.0 and cheap. The permissive licence is the point: you can use it commercially with no strings, self-hosted or through the API, which is rare at this quality level.

    • 256k context
    • $0.15 / $0.6 in / out per 1M tokens
    • open weights

    Still the Mistral 256k ceiling. On the Mistral API batch processing is half price and cached input is 90% cheaper, which is the most aggressive caching discount in this catalogue.

    Weights are Apache 2.0, so downloading and self-hosting costs nothing but hardware.

    Ministral 3 8B

    Mistral

    Laptop-sized and Apache 2.0, with the unusual property that input and output cost exactly the same — which makes it the easiest model in the catalogue to budget for.

    • $0.15 / $0.15 in / out per 1M tokens
    • open weights

    The family is 3B at $0.10 / $0.10, 8B at $0.15 / $0.15 and 14B at $0.20 / $0.20. Use small models for one narrow job at a time. An agent loop needs 64k or more of context and dependable tool calling, and that is exactly where models this size start to wobble.

    Weights are Apache 2.0 and the sizes are small enough for a normal laptop.

    Qwen3.8 (open weights, 27B)

    Alibaba (Qwen)

    The open-weight Qwen line is the most-pulled non-Western family on Ollama, and 27B is the size a strong laptop or a single decent GPU can actually hold.

    • open weights

    The older Qwen3.6 (27B and 35B) still has 6.7M pulls against 2.3M for Qwen3.8, so the community default is one generation behind — worth checking which one your tutorial actually assumes. The Ollama guide for running Claude Code against a local model uses a Qwen model in its own example.

    Free to download and run via Ollama.

    GLM-5.3

    GLM

    The highest-scoring open-weight model on the independent Artificial Analysis Intelligence Index: 45 at max effort on v4.3.2, against 53 for Claude Fable 5.1 and GPT-6 Astra.

    • open weights

    Two teaching points. First, the open-weight gap behind the frontier is visible but modest, and 66 of the 149 models the index evaluates are open-weight. Second, on Ollama this one exists only as a cloud model — open weights never meant it fits on your machine. Scores depend on the effort setting, so always name the setting when you quote one.

    Kimi K3

    Kimi

    The second-strongest open-weight score on the Artificial Analysis index (44 at max effort), and available as a model choice inside Cursor alongside Claude, GPT and Gemini.

    • open weights

    Cloud-scale like GLM-5.3: Ollama carries the Kimi line only as cloud models. The practical lesson from both is that the open-weight frontier and the runs-on-my-laptop frontier are two completely different lines, roughly 500 billion parameters apart.

    On pricing

    Every hosted model runs two meters. Tokens in — your prompt, your files, every tool result, and the entire conversation so far — and tokens out, which is what it writes back. Prices on this page are USD per million tokens at the standard non-batch, non-cached rate. Output is the expensive meter everywhere: 4 to 6 times input across every vendor here, because input is read in a single pass while output has to be produced one token at a time. You might conclude that output dominates your bill, and for a chatbot it does. For a coding agent it does not, because the agent re-sends its whole context on every single turn of the loop, so a long session is mostly input, read over and over. That is why the discounts that actually matter are on the input side: cached input costs 10% of the normal rate at Anthropic and OpenAI (2.5% on Fable 5.1) and 90% less at Mistral, and batch processing is roughly half price everywhere. A concrete size to hold in your head: 100k input plus 10k output, about what it takes to read a mid-size codebase and write one feature, costs around $1.50 on Fable 5.1, $0.75 on Opus 5, $0.30 on Sonnet 5, $0.11 on Gemini 3.8 Flash and $0.032 on GPT-5.6 Luna. One last thing that hits us specifically: the same sentence in Kazakh costs roughly twice its English token count on the best modern tokenizers, and up to 4.7 times on Llama-family ones. Write your prompts, your code comments and your CLAUDE.md in English, and keep Kazakh for what your users actually read.

    05 Running it on your own machine

    OllamaThe standard way to run an open-weight model on your own machine: one command pulls a model, another chats with it, and it exposes a local server other tools can talk to. Nothing you type leaves your laptop unless you deliberately choose a cloud model. Use ollama ps to see how much of the model ended up on the GPU versus the CPU.A terminal, and honest arithmetic about memory: roughly 0.6 to 0.7 GB per billion parameters for a 4-bit model, plus 1 to 2 GB for context. The anchor to remember is gpt-oss:20b — a 14 GB download that Ollama says runs with as little as 16 GB. Change the default context length before anything else: it ships at 4,096 tokens (OLLAMA_CONTEXT_LENGTH), and agent work needs 64k or more. On Windows the models land in C:\Users\<you>\.ollama\models, so check you have the disk space.LM StudioThe same job in a desktop app: a model browser, a chat window and a local server, with no terminal involved. The gentler first step if the command line is still new, and a good way to watch memory use climb as you load a bigger model.macOS: Apple Silicon only (M1 to M4), macOS 14.0 or newer, 16 GB or more of RAM recommended. Windows: x64 with AVX2, or a Snapdragon X Elite ARM machine; at least 16 GB of RAM and at least 4 GB of dedicated VRAM recommended. Linux: Ubuntu 20.04 or newer. The old claim that 8 GB Macs work fine with small models is not on the current requirements page — treat it as untested.Sizing: what your laptop actually fits8 GB of RAM: models of 1 to 4 billion parameters. Fine for chatting, summarising and explaining code; weak at multi-file agent work. 16 GB: 7 to 14 billion, or gpt-oss:20b with everything else closed. 16 GB or more of Apple Silicon, or an NVIDIA GPU with 8 GB or more: usable 12 to 30 billion parameter models — the first tier where a local coding agent becomes bearable rather than a demo.Treat this as a rule of thumb, not a promise. Speed on CPU alone is typically in the single digits to low teens of tokens per second, but that figure is unverified and depends heavily on your hardware. Measure on your own machine before you plan a workshop or a deadline around it.Claude Code pointed at a local modelThe answer to the question everyone asks: yes, technically you can run Claude Code against a model on your own machine. Ollama documents it — run ollama launch claude, or set ANTHROPIC_AUTH_TOKEN to ollama, ANTHROPIC_API_KEY to an empty string and ANTHROPIC_BASE_URL to http://localhost:11434, then start claude with the --model flag.Raise the context length to 64k or higher first — the Ollama docs say so explicitly for larger repositories — and expect a real drop in quality. An agent loop needs long context and dependable tool calling, and on an 8 to 16 GB laptop you only fit small models, which are weak at exactly those two things. Kazakh output suffers twice over: the tokenizer tax plus very little Kazakh in a small model.Ollama CloudThe escape hatch for when the model you want does not fit: the same Ollama commands, executed on someone else GPU. Not local and not private, but it is how you reach glm-5.3, deepseek-v4-pro or kimi-k2.6, which exist on Ollama only as cloud models.An account. The free plan has starter usage credits and 1 concurrent request; Pro is $20 a month with $60 of usage credits and 3 concurrent requests; Max is $100 a month with $300 of credits and 10 concurrent requests. Be clear with yourself that the privacy argument for running locally disappears here.
    Learn how to actually use these →