Coding models compared
The main AI models for writing code, side by side, with the date of every number. Use it to shortlist two models, then test them on your own work.
Version 1.0 ·
Prices are in US dollars per million tokens, input / output, at the standard rate for short prompts, as listed by each company on October 28, 2026. A token is about three quarters of an English word. “Context” is how much text the model can read at once. Scores marked “vendor” come from the company that makes the model; other scores come from independent testers.
Hosted models (you use them through an app or an API)
| Model | Maker | Released | API price (in / out) | Context | Coding tool from the same maker |
|---|---|---|---|---|---|
| Claude Opus 5.5 | Anthropic | Sept. 22, 2026 | $4 / $20 | 1M tokens | Claude Code |
| Claude Sonnet 5.5 | Anthropic | Sept. 28, 2026 | $2 / $10 | 1M tokens | Claude Code |
| Claude Haiku 5.5 | Anthropic | Oct. 7, 2026 | $0.10 / $0.50 (prompts up to 100K tokens) | 1M tokens | Claude Code (helper tasks) |
| Claude Fable 5.1 | Anthropic | Sept. 1, 2026 | $10 / $50 | 1M tokens | Claude Code |
| GPT-6 Astra | OpenAI | Sept. 2026 | $10 / $50 | not listed | Codex |
| GPT-6.1 Sol | OpenAI | Sept. 2026 | $2 / $10 (prompts up to 272K tokens) | 1.05M tokens | Codex |
| GPT-6.1 Terra | OpenAI | Oct. 2026 | $1 / $5 | not listed | Codex |
| GPT-6 Luna | OpenAI | Sept. 2026 | $0.10 / $0.50 | not listed | Codex |
| Gemini 4 Argon | 2026 | $2 / $10 until Nov. 30, 2026, then $4 / $20 | about 1M tokens | Antigravity | |
| Muse Wonder | Meta | Oct. 12, 2026 | $4 / $20 | 1M tokens | Muse Code |
| Muse Spark 1.3 | Meta | Sept. 2, 2026 | not listed | not listed | Muse Code |
| Grok 4.7 | SpaceXAI (formerly xAI) | Sept. 21, 2026 | $2 / $6 (prompts under 200K tokens) | 500K tokens | Grok Build |
Two price pairs line up exactly: Claude Sonnet 5.5 and GPT-6.1 Sol both cost $2 / $10, and Claude Haiku 5.5 and GPT-6 Luna both cost $0.10 / $0.50. Batch processing (results within a day instead of right away) costs about half on both platforms.
Open-weight models (you can download and run them)
| Model | Maker | Licence | Size | What it takes to run |
|---|---|---|---|---|
| Qwen3.8-27B | Alibaba (Qwen) | Apache 2.0 | 27 billion parameters | About 18 GB to download in its 4-bit version (56 GB at full precision): a 24 GB graphics card, or a computer with plenty of unified memory such as a recent Mac |
| Muse Glimmer-30B | Meta | Apache 2.0 | about 30 billion parameters | About 18 GB in its 4-bit version: the same kind of machine as Qwen3.8-27B |
| Muse Spark Open | Meta | Muse Model License (custom) | 816 billion parameters (about 42 billion active) | Data-centre servers |
“Apache 2.0” lets you use, change and sell what you build, with few conditions. A custom licence can add limits: Meta’s Muse Model License, for example, asks products to display “Built with Muse Spark Open” and requires a separate agreement for companies with more than $20 million in yearly revenue that sell access to the model. “Active” parameters are the part of the model that works on each word; it is why a very large model can still answer at a reasonable speed.
Coding scores: read them with care
Each lab now picks its own tests, and the tests come in several versions (Terminal-Bench 2.0, 2.1 and 4.0 are different exams). Compare numbers only inside the same row of the same test.
| Test | Results |
|---|---|
| Terminal-Bench 4.0 (vendor) | Claude Sonnet 5.5: 70.6% · Claude Opus 5.5: 66.4% · Claude Fable 5.1: 55.8% · Claude Haiku 5.5: 39.2% · Grok 4.7: 37.6% |
| Terminal-Bench 4.0 (Vals AI, independent) | GPT-6.1 Sol: 55.05% · Muse Spark Open: 6.6% |
| Android Bench (Google, results of October 26, 2026) | Gemini 4 Argon: 21.3%, 6th place; the ranges of error are wide, so close scores are ties |
| SWE-bench Pro (vendor, model card) | Qwen3.8-27B: 61.7 |
| SWE-bench Verified and Pro (vendor, model card) | Muse Glimmer-30B: 76.0 and 51.2 |
The big labs no longer report SWE-bench Verified for their newest models, and the public leaderboards that used to compare everyone have not added them. That is why your own test, on a task from your own work, tells you more than any table.
Claude Code and Codex at a glance
| Claude Code (Anthropic) | Codex (OpenAI) | |
|---|---|---|
| Free plan | Not included in Claude Free | Included in ChatGPT Free, with GPT-6 Luna |
| Cheapest paid plan | Claude Pro, $20 a month ($17 billed yearly) | ChatGPT Go, $8 a month (light tasks); Plus, $20 a month |
| Heavy use | Claude Max, from $100 a month | ChatGPT Pro, $100 to $500 a month |
| How limits work | A rolling 5-hour window plus a weekly limit | A rolling 5-hour window, weekly limits may apply |
| Where it runs | Terminal, VS Code, JetBrains, desktop app, web, mobile | Terminal, IDE extension, ChatGPT desktop app, web, mobile |
| Sandbox | Off by default; turn it on with /sandbox | On by default (workspace-only writes) |
| Use your own API key | Yes | Yes, without the cloud features |
Which one to start with
| Your situation | A good first pick |
|---|---|
| Learning, no budget | Codex on ChatGPT Free, or a free open model such as Qwen3.8-27B through Ollama or LM Studio |
| Private or offline work | Qwen3.8-27B or Muse Glimmer-30B on your own computer |
| First real projects, $20 a month | Claude Pro with Claude Code, or ChatGPT Plus with Codex: pick the one whose chat app you already use |
| Building your own app on an API | A mid-priced model (Claude Sonnet 5.5 or GPT-6.1 Sol), with Claude Haiku 5.5 or GPT-6 Luna for bulk work |
| Hard problems the mid-tier model can’t solve | Claude Opus 5.5 or GPT-6 Astra, for those tasks only |
Sources
- Anthropic: Claude models overview and Claude pricing
- Anthropic: Introducing Claude Opus 5.5, Sonnet 5.5, Haiku 5.5
- Claude Code documentation and sandboxing
- OpenAI API pricing and GPT-6.1 Sol model page
- OpenAI Codex models, pricing and limits and sandboxing
- Google Gemini API pricing and models
- Meta: Introducing Muse Wonder, Muse Spark 1.3, Muse Glimmer-30B model card
- SpaceXAI: Grok 4.7 and models and pricing
- Qwen3.8-27B model card and Ollama: Qwen3.8
- Google: Android Bench
- Vals AI: GPT-6.1 Sol and Muse Spark Open