RU EN
Join the waitlistWaitlist
ProgrammePricesAgentsBlog Sign in Русская версия

Your team's AI agent: which coding agent fits

Fourteen coding and office agents in one table. Answer three questions – who sits down at the tool, where company data can live, and whether the team works from Russia – and the table keeps the ones that fit. Every agent has its own page: the price, whether it is free, whether it works from Russia, and how to check its work. Data and prices come from the review "Coding Agents 2026" and are updated along with it.

Who works with it
Where company data may live
The team works from Russia

All 14

Tool For Free Data Russia Code
Kilo CodeKilo / AnacondaTerminal, IDE, desktop and cloud in one program; open, free clients, $15 per seat for a team, models at the provider's rates with no markup. Code and terminal clients are free; Teams $15 per seat program – yes cloudown key Russia – yes open
ZCodeZ.aiThe official GLM-5.3 harness, open source under Apache 2.0; the GLM Coding Plan meters token-based credits per 5 hours and per week. Code and terminal Lite $18 / Pro $80 / Max $168 a month; 30% off yearly no cloud Russia – yes open
Kimi Code CLIMoonshotAn open MIT shell for the terminal, desktop and VS Code; Kimi Code comes with a Kimi membership from Plus up, and Moonshot doesn't restrict access from Russia. Code and terminal CLI is free; model by Moonshot key or a Kimi plan from ¥99 (Plus) program – yes cloudown key Russia – yes open
DeepSeek HarnessDeepSeekA layer on top of other agents, not a replacement: plugins, a log of every step, and Claude Code and Codex called as subagents. Code and terminal your own key or a DeepSeek account, pay per token program – yes own key Russia – yes open
OpenCodeсообществоMIT license, no vendor account involved; with a local model, data never leaves the server, and teams get a console with budgets, SSO and the Go Teams plan. Code and terminal your own key, OpenCode Go $10 a month or Go Teams $20 per person program – yes own keymachine Russia – yes open
Kimi WorkMoonshotRuns on the user's machine: mounts folders, drives a logged-in browser, runs on a schedule. Documents and spreadsheets Free with limits; Plus $19 (Agent Swarm), Pro $39 (Goal mode); Business $600 per seat a year with limits machine Russia – yes closed
Tencent WorkBuddyTencentSplits a task into pieces and hands back finished reports, decks, and spreadsheets for review. Documents and spreadsheets Free 100 credits / Pro $10 / Team $40 per seat with limits cloud unconfirmed closed
Grok BuildxAIA bet on parallelism: subagents in isolated worktrees, background workflows and a headless mode for CI; included in every paid X Subscription plan. Code and terminal free with limits / Plus $8 / Premium $30 / Super $100 / Ultra $200 with limits cloud Russia – no open
GitHub Copilot CLIMicrosoftModels from Anthropic, OpenAI, Google and Microsoft in one subscription; Copilot Fusion picks the model, and a task moves between CLI, web and phone. Code and terminal Free / Pro $10 / Pro+ $39 / Max $100, usage in AI credits with limits cloud Russia – no closed
Antigravity CLIGoogleReplaced Gemini CLI; Gemini 4 Argon, Claude 5.5 and GPT-OSS in one terminal, with subagents and scheduled tasks. Code and terminal free with limits / Pro ≈$20 ($19.99) / Ultra $99.99 or $199.99 with limits cloud Russia – no closed
Claude CodeAnthropicThe most mature subagent system; included in any paid Claude plan. Code and terminal Max $100 / Ultra $200 per seat no cloud Russia – no closed
Codex CLIOpenAIOpen source, about 133,000 stars; one agent across terminal, IDE, web and cloud, with scheduled tasks. Code and terminal Free and Go with GPT-6 Luna / Plus $20 / Pro $100–500 / your own key with limits cloudown key Russia – no open
Claude CoworkAnthropicDocuments, spreadsheets and decks without a terminal: you hand over the task in an ordinary Claude chat and the work runs in Anthropic's cloud, even with the laptop closed. Documents and spreadsheets included in Claude plan no cloud Russia – no closed
Copilot CoworkMicrosoftA cloud sandbox inside Microsoft 365: access is granted by an admin, actions land in the audit log. Documents and spreadsheets Microsoft 365 Copilot licence plus credits, 1 credit = 1 cent by default no cloud Russia – no closed

Models in our benchmark →

What it costs a team

The monthly bill for seats on the cheapest paid plan of each tool in the table. Pay-per-token tools are left out: their bill depends on how much work they do.

  • Kilo Code$75
  • ZCode$90
  • Grok Build$40
  • GitHub Copilot CLI$50
  • Antigravity CLI$100
  • Claude Code$500

a month. Kimi Code CLI, DeepSeek Harness, OpenCode, Kimi Work, Tencent WorkBuddy, Codex CLI, Claude Cowork, Copilot Cowork – pay per token.

Where agents go wrong and how to catch it

Eight practitioner write-ups from August and September plus Matt Pocock's series and three Claude-channel videos, seven recurring stories. These are observations, not measurements.

access
Default permissions are wide open: in Oh My Pi the approval mode is literally called YOLO, and the first thing people recommend changing is turning on confirmation prompts.
time
Caution fades over time: first one repo, then "trust builds up" and the restrictions disappear.
traces
An autonomous agent leaves behind files, scripts, and changed commands, and "over time wrecks the machine if nobody watches it."
plugins
Around DeepSeek Harness, about 165 plugins piled up in three weeks; some of the published ones are malicious.
sandbox
Matt Pocock: an unattended agent needs a sandbox, not "skip all permission prompts"; the two cases he names are a deleted home directory and data exfiltration.
auto-approval
In Claude Code auto mode a separate classifier judges each risky action, not the agent itself; by the vendor's figure people approve 97% of permission prompts anyway.
silent stop
On the long-horizon Vending-Bench test cheap models fail quietly, not loudly: they stop ordering stock. Delegated work is monitored for inaction, not for error logs.

A terminal agent sees the repo, keys in environment variables, and the commit history. Its blast radius is wider than a desktop agent's, and it passes through security review less often.

Picking an agent takes an evening. After that the question is not the price but whether you can trust what it did.

Stop re-checking AI by hand

Checking everything an agent does by hand throws away the gain. Not checking means made-up numbers in a report. We are preparing a short course: in a week you build a check for one of your recurring AI outputs – a second agent and a couple of automatic cross-checks find the errors before your manager does.

What the course covers: programme and price

What changed this year

127 changes to prices, access, models and licences since 25 February 2026. The latest ten:

  1. ZCodeZ.ai made its flagship GLM-5-Turbo model fully closed ↗
  2. SpaceX sharply cut Grok Bot's weekly usage limits for subscribers, citing unsustainable costs ↗
  3. Kimi Code CLIA Kimi subscription can be paid with a Russian Mir card and used without a VPN, according to a Russian developer's report on Habr ↗
  4. Claude CodeAnthropic renamed its plans: Max 5x became Max at $100 a month and Max 20x became Ultra at $200; verified Ultra accounts no longer have the five-hour limit ↗
  5. DeepSeek HarnessDeepSeek API web search and page fetch, including from Harness, Claude Code and Codex, now cost $0.75 per 1,000 successful requests ↗
  6. Codex CLIOpenAI disclosed Codex CLI and Desktop flaws that let a hostile repository run code without approval; CLI 0.131.0 or later fixes them ↗
  7. OpenCodeOpenCode 2.2.3 connects Claude through an installed Claude Code: the Claude Agent provider uses its login and models ↗
  8. Mistral released open weights for the trillion-parameter Mistral Large 4; in the API it costs $0.70 and $2.10 per million tokens ↗
  9. Codex CLIChatGPT Pro and Business subscribers can request up to $500 a month in API credits
  10. Mistral's Le Chonk model became free for all users ↗
Show all 127 changes
  1. Copilot CoworkMicrosoft: from December 1 new Copilot Business licences bought through resellers have usage billing on by default, capped at 4,000 credits (about $40) per user a month ↗
  2. Codex CLIOpenAI doubled the Work and Codex allowance on Pro $100 and Pro $500 and cancelled the planned Pro $200 cut ↗
  3. Tencent WorkBuddyTencent unveiled WorkBuddy's enterprise edition for international companies: SSO, per-department security policies, audit log ↗
  4. Kimi Code CLIKimi K3.1 Pro released as an open model: available in the Kimi apps and the API, full weights promised by 4 November ↗
  5. Google introduced Gemini 4 Carbon at $2 and $10 per million tokens; for now the model is available only to cyber defenders ↗
  6. Kilo CodeKilo added NVIDIA models: the multimodal Nemotron 3.5 Super VL and the free Nemotron 3 Super ↗
  7. GitHub Copilot CLIGitHub Enterprise Cloud now includes Copilot access with no bundled credits: usage is metered from the first request; self-serve comes later ↗
  8. Copilot CoworkCopilot Cowork added GPT 6.1 Sol and xAI's Grok 4.7, plus projects that keep one initiative's tasks, instructions and files together ↗
  9. GitHub Copilot CLIGitHub Copilot Automations can now be capped in AI credits per run, day, week or month; shared automations are billed to the repository owner ↗
  10. GitHub Copilot CLICopilot Fusion is generally available on all plans, including Copilot CLI: it routes between models, with a temporary 30% discount on model costs ↗
  11. GitHub Copilot CLICopilot cloud sessions run in GitHub sandboxes; from November 2026 their compute is billed against the GitHub Actions entitlement ↗
  12. Kilo CodeAnaconda, Kilo's owner, joined the Linux Foundation-led Open Secure AI Alliance to build shared open tools for securing agents ↗
  13. OpenCodeOpenCode launched Go Teams: $20 per member a month, each with their own limits on open models, one company bill ↗
  14. Kimi Code CLIKimi Code 2.4 imports Claude Code and Codex sessions from the last 15 or 30 days, and plans no longer self-approve in auto mode ↗
  15. Kilo CodeKilo released a desktop app (beta) and the Kilo Gateway, a single API gateway that accepts your own provider keys ↗
  16. Kilo CodeKilo added sign-in with ChatGPT, letting the ChatGPT subscription pay for usage across its cloud surfaces ↗
  17. GitHub Copilot CLIGPT-6.1 Terra arrived in GitHub Copilot, a cheaper model for agentic coding and terminal work ↗
  18. Codex CLIGPT-6.1 Terra arrived in Codex on Plus, Pro, Business, Enterprise and Edu ↗
  19. Kimi WorkKimi Work added enterprise workspaces, company-built plugins and IP-allowlist access blocking ↗
  20. Antigravity CLIOpenAI models and Codex became available on Google Cloud, including in Antigravity ↗
  21. DeepSeek open-sourced its production inference engine DeepInfer under the Apache 2.0 licence ↗
  22. Anthropic opened a research preview of the Claude Duet agent to Pro and Max subscribers in the US, via a waitlist ↗
  23. Kilo CodeKilo offers Cursor enterprise customers Kilo Enterprise free for the rest of their Cursor contract, up to 12 months; apply by 30 November, model usage billed separately ↗
  24. The US ordered federal civilian agencies to drop DeepSeek, including third-party products built on it ↗
  25. Antigravity CLIGemini 3.8 Flash launched with introductory pricing until year end: $0.75 per million input tokens and $3.75 output ↗
  26. Meta released the open Muse Glimmer 30B model for local use ↗
  27. Antigravity CLIClaude Opus 5.5 and Sonnet 5.5 became available in Antigravity for Google Cloud customers once an admin enables them ↗
  28. Antigravity CLIGemini 4 Argon API pricing doubles on 1 December: $4 per million input tokens and $20 output instead of $2 and $10 ↗
  29. Kilo CodeKilo Code launched iOS and Android apps: cloud agents and remote control of the CLI from a phone ↗
  30. Kilo CodeKilo added Auto Model routing with three tiers: frontier, efficient and free models ↗
  31. GitHub Copilot CLIGitHub Copilot added a REST API for enterprise organisations to manage model policies centrally ↗
  32. MiniMax released M3.1 Flash: $0.15 per million input tokens and $0.60 per million output, 70% cheaper than the previous version ↗
  33. Inception released Mercury 3, a diffusion language model: up to a million tokens of context, $0.20 and $0.70 per million tokens ↗
  34. DeepSeek HarnessDeepSeek-V4.1-Flash-1017 released with vision, tuned for DeepSeek Harness; new Search and Fetch APIs ↗
  35. Replit Agent now defaults to OpenAI models; Claude can still be selected ↗
  36. ZCodeZ.ai split its keys: a standard API key bills the balance, only a Coding Plan key draws on plan quota ↗
  37. GitHub Copilot CLIOn 16 November GitHub Copilot will retire Claude Opus 4.8, Opus 5 and GPT-5.6 Terra; replacements are Opus 5.5, GPT-6.1 Sol or GPT-6 Luna ↗
  38. GitHub Copilot CLIGemini 4 Argon is generally available in GitHub Copilot on Pro+, Max, Business and Enterprise ↗
  39. Antigravity CLIGemini 4 Argon arrived in Antigravity CLI and app for Google AI Pro and Ultra subscribers ↗
  40. Kilo CodeKilo released the Kilo MCP server: an agent can report spend and start cloud agents straight from a chat ↗
  41. Cursor is now sold as an X Subscription: the $8 Plus plan offers only Grok models, higher plans bill third-party models separately at API list price ↗
  42. Grok BuildGrok Build is now part of the X Subscription: Plus $8, Premium $30, Super $100, Ultra $200 a month; limited on the free plan ↗
  43. OpenCodeOpenCode Console added policies: monthly budgets and rules for who may use which models, providers and tools ↗
  44. Claude CodeClaude Haiku 5.5: $0.10 and $0.50 per million tokens for prompts up to 100k – about 75% cheaper to run than Haiku 4.5 per Anthropic; Sonnet 5.5 cache reads halved ↗
  45. Claude CoworkNew Cowork tasks on Pro, Max and Ultra run only in Anthropic's cloud: schedules run with the computer off, local-only mode was removed ↗
  46. Kimi WorkKimi Work added human–AI co-editing of Notion and Lark documents ↗
  47. Codex CLIOpenAI released GPT-6.1 Sol: $2 per million input tokens and $10 per million output, five reasoning effort levels at one price ↗
  48. Codex CLIChatGPT Pro is now three tiers – $100, $200 and $500 a month; Pro includes always-on Dots agents ↗
  49. Copilot CoworkCopilot Cowork uses usage-based billing: the bill scales with the work you delegate ↗
  50. GitHub Copilot CLIClaude Sonnet 5.5 is generally available in GitHub Copilot – aimed at well-scoped everyday work like building features and fixing bugs ↗
  51. Codex CLIOpenAI cancelled the release of Astra 6.1 after it failed internal safety testing ↗
  52. Claude CodeBugHunt: Sonnet 5.5 at max effort fixed the most bugs – 55.5 of 105 on average against 41.7 for Opus 5.5 – but a run cost about $135 against $59 and took almost four hours against 67 minutes: for every fix it writes a test and confirms the test fails with the fix reverted ↗
  53. Kilo CodeKilo Code integrated Enkrypt AI guardrails: admins define policies once and developers keep working as usual ↗
  54. Freebuff: where the full mode is unavailable, a Limited mode runs DeepSeek V4 Flash with six free one-hour sessions a day ↗
  55. Codex CLIOpenAI is preparing a $500-a-month ChatGPT Pro Max plan, according to TestingCatalog; no official announcement yet ↗
  56. ZCodeZCode's source code was opened on GitHub under the Apache 2.0 licence ↗
  57. Codex CLIOpenAI released GPT-6 Sol and GPT-6 Luna: API prices half of the GPT-5.6 promo rates (Sol at $2 and $10 per million tokens) ↗
  58. Kilo CodeKilo Swarm is on by default: agents and subagents exchange messages on a shared board; the vendor measures a 39% cost cut on the tasks it is built for and inconclusive results elsewhere ↗
  59. Kilo CodeKilo Code can import global Claude Code settings once (an experimental feature, off by default) ↗
  60. Antigravity CLIThe Antigravity SDK now runs local models offline, starting with Gemma 4 26B via LiteRT ↗
  61. Kimi Code CLIFireworks released Ember-1, a Kimi K3 derivative that matches its quality with 40% fewer tokens ↗
  62. Claude CodeClaude Opus 5.5 released: $4 per million input and $20 per million output tokens – 20% cheaper than Opus 5, with cache reads 60% cheaper ↗
  63. Codex CLICodex CLI 0.157.0 added GPT-6 Sol and GPT-6 Luna, including Amazon Bedrock support and migration prompts for older models ↗
  64. Grok BuildGrok 4.7 released: $2 per million input and $6 per million output tokens; a fast variant is twice as fast at twice the price ↗
  65. DeepSeek HarnessDeepSeek API: Chinese national holidays now count as discounted off-peak days, on top of weekends ↗
  66. ZCodeTraffic analysis: while a user is logged in, ZCode silently uploads the workspace, including the full git history, to Alibaba Cloud storage ↗
  67. Kimi Code CLIKimi K3 arrived on Amazon Bedrock – the first open-weights model there with explicit prompt caching ↗
  68. Claude CodeRedesigned Projects in Claude Code entered beta for select Pro and Max subscribers on cloud sessions; others can join the waitlist ↗
  69. Kimi Code CLIKimi Code CLI 2.0: moved to Node.js with a desktop app; the old Python kimi-cli client is archived ↗
  70. Claude CoworkCowork was merged into the ordinary Claude chat: on Pro and Max any conversation can take a task, with shared projects and skills ↗
  71. Claude CodeAnthropic reworked Claude pricing: Pro $17 a month billed yearly, Max from $100, Team $20 / $100 per seat, Enterprise $20 per seat plus API usage ↗
  72. The Enkrypt AI product page confirms it is part of Anaconda; existing products, plans and support continue unchanged ↗
  73. GitHub Copilot CLIGitHub Copilot code review resolves its own comments after the feedback is addressed and suggests commit messages when its fixes are applied ↗
  74. Claude CodeClaude Code v2.1.269: a configurable limit of up to 256 concurrent agents per Workflow run; claude plugin eval runs plugin evaluations with scores and reports ↗
  75. Codex CLIOpenAI paused sales of the $200 Pro subscription because of Astra demand; no new Codex seats on that tier for now ↗
  76. Claude CodeClaude Code gained an Advisor mode: a second Opus checks the first one's work, the request context grew to 1M tokens ↗
  77. Kilo CodeKilo Code now shows an Enkrypt safety score for models on a 0–100 scale, lower is safer ↗
  78. DeepSeek HarnessDeepSeek V4.1 Flash released: new architecture, 763B parameters, open weights, 1M context, native multimodality; a radically smaller KV cache makes long agent sessions cheaper and faster, and the API is another 50% cheaper off-peak ↗
  79. Cursor released Projects: months-long context and a coordinator delegating to subagents; the vendor's internal data says new users merge 30% more pull requests ↗
  80. GitHub Copilot CLIGitHub Copilot app added Jira integration: issues enter a shared canvas with context for investigation, implementation and pull request preparation ↗
  81. GitHub Copilot CLIMAI-Code-1-Flash retired across GitHub Copilot; the suggested replacement is MAI-Code-1.1-Flash, which Enterprise admins may need to enable in model policies ↗
  82. OpenAI introduced ChatGPT for Financial Services on GPT-6 Astra with built-in data from LSEG News, PitchBook, Daloopa and Crunchbase ↗
  83. The Data Agent in ChatGPT Work connects to approved company sources, investigates metric changes and builds interactive dashboards; administrators control access ↗
  84. Codex CLICodex lead: ChatGPT Work is the full Codex harness in a cloud VM with internet access; the harness is open and not tied to OpenAI models ↗
  85. Kilo CodeKilo Code added Auto Model: the shell picks a model per request, the vendor claims about 30% savings on mid-size tasks ↗
  86. Claude Managed Agents: hosted agents for products; founders recommend a separate verifier agent with a clean context ↗
  87. Cursor closed to Russia: completion and the agent stopped for everyone, a personal API key does not help ↗
  88. Kilo CodeGPT-6 Astra and Claude Fable 5.1 are available in Kilo Code across VS Code, JetBrains and the CLI ↗
  89. Codex CLIOpenAI released GPT-6 Astra: the top tier in Codex, 2.5 times the price of GPT-5.6 Sol ↗
  90. Kilo CodeKilo Code is on the AWS Marketplace: $180 a year per user for teams, procurement without a new vendor ↗
  91. Claude CodeAnthropic released Claude Fable 5.1: available in Claude Code as the top-tier model at $10 / $50 per million tokens
  92. OpenAI will cut Cursor off from its models on 12 November 2026 after the ownership change ↗
  93. Claude CoworkClaude Cowork gained a built-in browser: it opens sites, fills forms and collects data on its own ↗
  94. Kimi Code CLIMoonshot open-sourced the Kimi K3 weights, 2.8 trillion parameters: the largest open-weights release to date ↗
  95. Claude CoworkClaude for Word: the agent edits a document with track changes on and checks numbers against the source ↗
  96. Cursor moved to hourly metered pricing without an announcement: users learned from an email while the docs still showed a flat price ↗
  97. Kilo CodeKilo Code released a mobile app: agent sessions, review and merging pull requests from a phone
  98. DeepSeek HarnessDeepSeek V4 Flash Vision, an experimental multimodal version: reads screenshots and diagrams, files up to 64 MB ↗
  99. GitHub Copilot CLIGitHub Copilot moved to metered token billing and dropped annual plans; a budget approved in spring is out of date ↗
  100. ZCodeGLM 5.3 shipped: 1M-token context, API at $1.4 / $4.4 per million; ZCode runs on it by default ↗
  101. SpaceX closed its $60B acquisition of Cursor ↗
  102. Claude CodeClaude Code switched to auto mode by default: a classifier judges commands instead of allow/deny prompts ↗
  103. DeepSeek HarnessDeepSeek Harness released: everything is a plugin, it calls Claude Code and Codex as sub-agents
  104. Claude CodeAnthropic released the Claude apps gateway for AWS: team access to Claude Code and Desktop via SSO, spend limits and policies ↗
  105. Grok BuildGrok Build left beta with version 1.0 under the Apache 2.0 licence
  106. Tencent WorkBuddyWorkBuddy refreshed its personal plans: Pro at $10 a month with 2,000 credits, the free plan at 100 credits a month ↗
  107. Claude CodeClaude Code auto mode explained: a separate classifier judges risky actions, admin rules cannot be removed ↗
  108. DeepSeek HarnessDeepSeek V4 Flash 0731: the update beat V4 Pro on benchmarks with 10 times fewer active parameters, open weights, $0.14 per million tokens ↗
  109. Claude CodeClaude Code now defaults to Opus 5; Fable 5 and Sonnet 5 are switched manually
  110. Claude CodeThe Claude Code system prompt was cut by about 80% for frontier models
  111. Kilo CodeKilo Code acquired by Anaconda: about 10 trillion tokens a month pass through the platform
  112. Claude Tag: an agent inside Slack that takes a standing objective per channel ↗
  113. Claude CoworkClaude Cowork: delegation and scheduled tasks; in June they ran only while the desktop app was open ↗
  114. Antigravity CLIGemini CLI was shut down and replaced by Antigravity CLI
  115. DeepSeek HarnessDeepSeek released V4 Pro and V4 Flash: analysing 10 thousand lines of code in Flash costs under a cent ↗
  116. Qwen Code closed its free OAuth access: a key is now required
  117. GitHub Copilot CLIGitHub Copilot CLI opened to every Copilot subscriber

How companies rolled them out

Stories from the official Claude channel (June to September 2026), Revolut's talk at AI Engineer NYC (October 2026) and the Codex lead's interview on The Pragmatic Engineer: told by the companies and the vendors, figures not verified externally.

Revolut
About 2,000 of the bank's engineers work on the open GLM model, deployed on 56 dedicated GPUs and used from OpenCode and the Codex app. The reason is money: with Claude Code some engineers spent 1,000 to 10,000 dollars a month, caps made people ration tokens, and now the bill is flat and, by the company's account, a tenth of what it was. Claude Code and Codex on cloud models stay as exceptions; commands such as a build run only after a human approves them.
DoorDash
Every engineering manager had to ship real production code through all the company's gates with an agent, to feel both the capability and the friction. A separate fast procurement and security path for AI tools.
DoorDash
A written write-up is required for wins and for failed experiments; one tech lead put about 50 architectural principles into markdown files in the repo. The bottleneck moved to product and design reviews.
Spotify
An LLM judge lifted automated PR success from roughly 20 to 30 percent to about 80; it was later removed once models caught up. Auto-merge forced teams to accept that not every change gets read, and to invest in test automation.

From the blog

Take the course in English – first

We’ll email you the day the English course opens. One email, nothing else unless you ask for it.