# AI model comparison for coding agents

URL: https://developertoolkit.ai/en/ecosystem/models/
Description: Coding-agent models, prices and defaults for Claude Code, Codex and Cursor, checked 2026-10-12, and the rule for choosing: default, then effort, then evals.

On 2026-10-12, Claude Opus 5.5 is the Claude Code default on both release channels and GPT-6.1 Sol is the Codex default, with GPT-6 Astra chosen by hand for the hardest work. The rule: start on the default, raise effort before switching models, and switch only when evals show a gain: Claude Fable 5.1 for the longest-horizon work; Claude Sonnet 5.5, GPT-6.1 Sol, Claude Haiku 5.5 or GPT-6 Luna when cost dominates.

In September 2026 Anthropic shipped Claude Fable 5.1 and Opus 5.5, OpenAI shipped GPT-6 Astra, Sol and Luna, Google shipped Gemini 3.8 Flash, and xAI shipped Grok 4.7. Since then Anthropic has added Claude Sonnet 5.5 (2026-09-28) and Claude Haiku 5.5 (2026-10-07), OpenAI has added GPT-6.1 Sol (2026-09-29), and Meta has added Muse Wonder (2026-10-12). Your model picker changed under you, a launch chart says your current model is behind, and someone on the team wants to switch today. This page is for the **developer** choosing a model for a session and the **tech lead** or **CTO** who sets the default and budget for a team. It is also the one page on this site that carries model versions and prices; every other page links here.

## What this page gives you

- A job-by-job model recommendation for Claude Code, Codex and Cursor.
- Current prices, model IDs and context limits, each checked on 2026-10-12.
- The exact commands and settings to change model and effort, and to pin them.
- Two copy-paste prompts: one finds stale model IDs in a repository, one runs an effort sweep.
- The failure modes that show up after a model change, with the fix for each.

## Which model should you use for each job?

Anthropic's own advice is: "If you're unsure which model to use, start with Claude Opus 5.5 for most workloads." OpenAI's Codex models page says: "For complex coding and agentic workflows, use GPT-6.1 Sol when available to your account and client." The table follows both vendors' defaults and positioning.

| Job | Claude Code | Codex | Why |
| --- | --- | --- | --- |
| Hard agentic coding (the default case) | Claude Opus 5.5 at default `medium` effort, raised to `high` or `xhigh` when work does not land | GPT-6.1 Sol (default effort `low`); GPT-6 Astra chosen by hand for the hardest work | Each tool's default; Anthropic's "start with" model; OpenAI: "Choose **Astra** when a task needs the strongest capability across steps and tools" |
| Everyday work | Claude Opus 5.5 at default effort, or Claude Sonnet 5.5 when cost matters | GPT-6.1 Sol ("Latest workhorse model for coding and everyday work") | Anthropic: Sonnet 5.5 "is strongest at well-scoped everyday tasks", while "Opus 5.5 remains clearly stronger at complex, open-ended work requiring sustained judgment". Sonnet 5.5 and GPT-6.1 Sol cost $2 / $10, half or less of the flagship price |
| Cheap, high volume: classification, extraction, subagent fan-out | Claude Haiku 5.5 (two-tier price: $0.10 / $0.50 up to 100K-token prompts, $0.50 / $2.50 above) | GPT-6 Luna ("Fast and affordable model for easier tasks") | Anthropic: Haiku 5.5 "pairs well with Opus 5.5 and Sonnet 5.5 as a subagent on coding work" and is best suited to "more narrowly scoped tasks … like compaction, summarization, or subagent work". OpenAI: "Use Luna for focused, repeatable tasks" |
| Longest horizon, hardest reasoning | Claude Fable 5.1, chosen by hand (`/model fable`) | GPT-6 Astra at `max` or `ultra` | Anthropic: use Fable 5.1 "when your evals on Claude Opus 5.5 at higher effort still fall short" |
| Self-hosted or air-gapped | See [open-weight models for coding agents](/en/ecosystem/models/open-weight/) | Same, via `codex --oss` | Licence and harness support differ per model |

**Cursor.** Cursor names no single default model: Hobby runs Agent, Chat and Tab with Auto, and no Cursor page states a default for paid plans (checked 2026-10-12). Cursor calls Grok 4.7 its flagship model ("Built for the hardest, longest-running tasks"), jointly trained by Cursor and xAI. Claude Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1 are third-party models billed at their API rates, and no GPT-6 model is in Cursor's model list (checked 2026-10-12). The Cursor tab in the next section has the rates.

## What do the current models cost?

Prices are USD per million tokens (MTok), input / output, standard tier, checked 2026-10-12 on each vendor's price page unless the row names another source.

**Anthropic**

| Model | API ID | Price in / out | Context / max output | Status |
| --- | --- | --- | --- | --- |
| **Claude Opus 5.5** | `claude-opus-5-5` | $4 / $20 | 1M / 128K | Current default, released 2026-09-22 |
| **Claude Fable 5.1** | `claude-fable-5-1` | $10 / $50 | 1M / 128K | Current; never a default |
| **Claude Sonnet 5.5** | `claude-sonnet-5-5` | $2 / $10 | 1M / 128K | Current, released 2026-09-28; the `sonnet` alias on the Anthropic API from Claude Code v2.1.284 |
| **Claude Haiku 5.5** | `claude-haiku-5-5` | $0.10 / $0.50 up to 100K-token prompts; $0.50 / $2.50 above | 1M / 128K | Current Haiku, released 2026-10-07; the `haiku` alias on the Anthropic API from Claude Code v2.1.293 (the `latest` channel) |
| Claude Sonnet 5 | `claude-sonnet-5` | $2 / $10, permanent | 1M / 128K | Legacy; the auto-mode classifier in Claude Code still runs on it by default |
| Claude Haiku 4.5 | `claude-haiku-4-5` | $1 / $5 | 200K / 64K | Legacy; no deprecation notice on 2026-10-12 |
| Claude Sonnet 4.5 | `claude-sonnet-4-5-20250929` | $3 / $15 | — | Deprecated 2026-09-30; retires 2026-11-30 on the Claude API, Claude Platform on AWS and Microsoft Foundry |
| Claude Opus 5 | `claude-opus-5` | $5 / $25 | 1M / 128K | Legacy, still served |
| Claude Opus 4.8 | `claude-opus-4-8` | $5 / $25 | — | Legacy, still served |

Batch processing halves every rate. Cache reads cost 0.1x input, except 0.05x on Opus 5.5 ($0.20) and Sonnet 5.5 ($0.10) and 0.025x on Fable 5.1 ($0.25). Sonnet 5.5's cache reads cost $0.20 at launch and $0.10 since 2026-10-07; Haiku 5.5's cost $0.01 up to 100K-token prompts and $0.05 above. Claude 4.6 and later models bill the full 1M context at the standard rate, except Haiku 5.5: a request's prompt length counts all its input tokens, including cache reads and cache writes, and a prompt over 100,000 tokens bills the whole request at $0.50 / $2.50. Opus 5.5 fast mode is $8 / $40, a research preview on the Claude API only. Claude 4.7 and later models use a tokenizer that produces about 30% more tokens for the same text, so compare cost per task, not price per token. Bedrock IDs add the `anthropic.` prefix, for example `anthropic.claude-opus-5-5`, `anthropic.claude-sonnet-5-5` and `anthropic.claude-haiku-5-5`.

The Sonnet 4.5 dates apply to Anthropic-operated platforms; Bedrock and Google Cloud set their own. Anthropic names `claude-sonnet-5-5` as the replacement. Claude Haiku 4.5 and Claude Opus 4.5 carry commitments of "not sooner than" 2026-10-15 and 2026-11-24, with at least 60 days' notice, and neither had a deprecation notice on 2026-10-12.

**OpenAI**

| Model | API ID | Price in / out (over 272K input) | Context (API) | Status |
| --- | --- | --- | --- | --- |
| **GPT-6.1 Sol** | `gpt-6.1-sol` | $2 / $10 ($4 / $15) | 1,050,000 / 128,000 | Current, released 2026-09-29; Codex bundled default since CLI 0.159.1 |
| **GPT-6 Astra** | `gpt-6-astra` | $10 / $50 ($20 / $75) | 1,050,000 / 128,000 | Released 2026-09-03; Codex bundled default 2026-09-04 to 2026-09-29, priority 2 since |
| **GPT-6 Sol** | `gpt-6-sol` | $2 / $10 ($4 / $15) | 1,050,000 / 128,000 | Released 2026-09-22; Codex label "Previous generation workhorse model." |
| **GPT-6 Luna** | `gpt-6-luna` | $0.10 / $0.50 ($0.20 / $0.75) | 1,050,000 / 128,000 | Current small model, released 2026-09-22 |
| GPT-5.6 Sol, Terra, Luna | `gpt-5.6-sol`, `gpt-5.6-terra`, `gpt-5.6-luna` | $4 / $20, $2 / $12, $0.20 / $1.20 | 1,050,000 / 128,000 (Sol) | Older; Sol's promotional price runs "at least through November 21, 2026" |
| GPT-5.5 | `gpt-5.5` | $5 / $30 ($10 / $45) | 1,050,000 / 128,000 | Retires from ChatGPT, ChatGPT Work and Codex on 2026-10-14; the OpenAI API is not affected |

Prices and API context come from OpenAI's pricing and model pages (read 2026-10-12). The maximum input is 922,000 tokens on every GPT-6 model. A request over 272K input tokens bills the whole request at the long-context rate in brackets. Cached input costs $0.10 on GPT-6.1 Sol, $1 on Astra, $0.20 on GPT-6 Sol and $0.01 on Luna. Batch and Flex are half the Standard rate, Fast mode is 2x and Ultrafast is 6x (GPT-6.1 Sol Ultrafast is $12 / $60). In Codex the context window is 272K by default and can be raised to 872K, according to the bundled model catalog in Codex 0.157.1 (checked 2026-09-26) and again in 0.162.1 (checked 2026-10-12). The OpenAI Python SDK's model list (version 3.29.0, released 2026-10-11, checked 2026-10-12) has no GPT-6 Terra, no `gpt-6.1-astra` and no `gpt-6.1-luna`. GPT-6 Pro is a ChatGPT product name, not an API model ID.

**Google, xAI and others**

| Model | API ID | Price in / out | Context | Status |
| --- | --- | --- | --- | --- |
| **Gemini 3.8 Flash** | `gemini-3.8-flash` | $0.75 / $3.75 through 2026-12-31, then $1.50 / $7.50 from 2027-01-01 | 1,048,576 in / 65,536 out | Current Flash, generally available 2026-09-02 |
| Gemini 3.1 Pro Preview | `gemini-3.1-pro-preview` | $2 / $12 up to 200K input, $4 / $18 above | 1,048,576 in / 65,536 out | Google still lists it as a preview |
| Gemini 4 Argon | none yet | Announced: $2 / $10 introductory, then $4 / $20 | Output limit 1M tokens | Announced 2026-09-30 for trusted cyber defenders only; no public API |
| **Grok 4.7** | `grok-4.7` | $2 / $6 up to 200K prompt tokens, $4 / $12 above | 500,000 | Newest Grok; xAI's release notes list it under September 2026 |
| Muse Wonder (Meta) | `muse-wonder-1.0` | $4 / $20 | 1M | Released 2026-10-12 on the Meta Model API and in Muse Code |
| Mistral Large 4 | `mistral-large-4` | $1.36 / $4.18 list; launch pricing is 50% off for two weeks | 1M | Public preview since 2026-10-06; open weights "coming soon" |

Google prices are global-endpoint prices from Google's Agent Platform pricing page (formerly Vertex AI; the old URL redirects), and the Gemini 3.8 Flash row matches the Gemini API's; regional endpoints cost 10% more. Google says Gemini 4 Argon is "rolling out to a set of trusted cyber defenders through our Fairwind Program", and it is not on the Gemini API models page or the price list on 2026-10-12, so there is no model ID to call yet; the price is the announced one. xAI calls Grok 4.7 "Our flagship model for code and everything else". Its cached input costs $0.50 ($1 above 200K), and the Grok 4.7 Fast variant is the same model at 2x the standard rates, "available only in Cursor and Grok Build" (xAI's release notes). Muse Wonder's cached input costs $0.20; Muse Wonder Full is available only to Muse Code subscribers on the Power Usage ($50 per month) and Maximum Usage ($200 per month) plans. Mistral Large 4 has 1.05T total and 52B active parameters, and its cached input costs $0.14.

**Cursor**

| Model (Cursor's spelling) | Input / cache write / cache read / output | Notes |
| --- | --- | --- |
| Grok 4.7 | $2 / — / $0.50 / $6; Fast $4 / — / $1 / $12 | First-party; 256K standard window, 500K long context; above 256K input, standard requests cost 2x and Fast requests 3x the standard rates |
| Composer 2.5 | $0.50 / — / $0.20 / $2.50; Fast $3 / — / $0.50 / $15 | First-party; Fast is the default in the product |
| Claude Opus 5.5 | $4 / $5 / $0.20 / $20 | Fast mode `claude-opus-5-5-fast` is $8 / $40 |
| Claude Sonnet 5.5 | $2 / $2.50 / $0.10 / $10 | Up to 1M context |
| Claude Haiku 5.5 | $0.10 / $0.125 / $0.01 / $0.50 | Up to 100K input; 5x above 100K |
| Claude Fable 5.1 | $10 / $12.50 / $0.25 / $50 | Data-retention opt-in |

Rates are from Cursor's model pages and pricing page (checked 2026-10-12). Cursor's Cursor Models pool is Grok 4.7, Grok 4.6, Grok 4.5 and Composer 2.5; the Claude models are Other Models, billed at the API price. Cursor says it charges "at these published API rates with no markup"; on Teams and Enterprise, third-party requests also carry a Cursor Token Rate of $0.25 per million tokens, and the first-party Grok and Composer models are exempt.

Tool subscription prices (Claude Pro and Max, ChatGPT plans, Cursor plans) live on [the tool pricing comparison](/en/comparison/pricing-analysis/), not here.

## How much does one task cost on each model?

A task that reads 200K uncached input tokens and writes 20K output tokens costs, at list price and before caching, 0.2 MTok of input at the model's input price plus 0.02 MTok of output at its output price:

| Model | Arithmetic | Cost per task |
| --- | --- | --- |
| Claude Fable 5.1 · GPT-6 Astra | 0.2 × $10 + 0.02 × $50 | $3.00 |
| Claude Opus 5.5 | 0.2 × $4 + 0.02 × $20 | $1.20 |
| Claude Sonnet 5.5 · Claude Sonnet 5 · GPT-6.1 Sol · GPT-6 Sol | 0.2 × $2 + 0.02 × $10 | $0.60 |
| Claude Haiku 4.5 | 0.2 × $1 + 0.02 × $5 | $0.30 |
| Claude Haiku 5.5 | 0.2 × $0.50 + 0.02 × $2.50 | $0.15 |
| GPT-6 Luna | 0.2 × $0.10 + 0.02 × $0.50 | $0.03 |

The Haiku 5.5 row uses the over-100K tier: a 200K-token prompt bills the whole request at $0.50 / $2.50, so the task costs $0.15, not the $0.03 that the up-to-100K prices would give. The 200K prompt stays under OpenAI's 272K threshold, so the OpenAI rows use the short-context prices. Token counts for the same text differ between vendors, so equal rows do not mean equal cost for the same task. Anthropic reports that, in its testing, Sonnet 5.5 "costs up to 30% less per task than its predecessor", a vendor claim that list-price arithmetic cannot show: the two have the same price per token.

This is arithmetic, not a measurement. Agents re-read context on every turn, so caching changes real cost more than the list price does, and a cheaper model that needs three attempts costs more than an expensive one that needs one. Anthropic's Claude Code cost documentation puts typical enterprise spend at "around $13 per developer per active day and $150-250 per developer per month" (checked 2026-09-26). Measure cost per *accepted* task on your own work before you budget.

## How do you change model and effort in each tool?

**Claude Code**

On 2026-10-12, Opus 5.5 is the default on every paid plan, the Anthropic API, Claude Platform on AWS, Bedrock and Google Cloud's Agent Platform, on both release channels: the default arrived in v2.1.280, and, on 2026-10-17, `stable` (v2.1.292) and `latest` (v2.1.302) are both past it. Microsoft Foundry still defaults to Sonnet 4.5, which Anthropic deprecated on 2026-09-30 and retires on 2026-11-30 (check the [deprecations page](https://platform.claude.com/docs/en/about-claude/model-deprecations) before relying on this default). Opus 5.5, Sonnet 5.5 and Haiku 5.5 run at `medium` effort by default in Claude Code; Opus 4.7 runs at `xhigh` and every other model that supports effort, Fable 5.1 and Sonnet 5 among them, at `high`. Sonnet 5.5's default on the API is `high`, so the same model gets a lower effort in Claude Code than in an API script unless one of them sets it. Haiku 4.5 has no effort setting.

```bash
# Terminal: start a session on a chosen model and effort
claude --model opus --effort high

# Try fallbacks in order when the primary is overloaded
claude --model opus --fallback-model sonnet
```

Inside a session, `/model` switches model and `/effort` sets `low`, `medium`, `high`, `xhigh` or `max`. Sonnet 5.5 needs Claude Code v2.1.284 or later and Haiku 5.5 needs v2.1.293 or later, which only the `latest` channel has on 2026-10-12. The aliases resolve per provider (Claude Code model configuration, read 2026-10-12), and Sonnet 4.5, the `sonnet` alias on Bedrock, Google Cloud and Foundry, retires on 2026-11-30 on Anthropic-operated platforms (Bedrock and Google Cloud set their own dates):

| Alias | Anthropic API | Claude Platform on AWS | Bedrock and Google Cloud's Agent Platform | Microsoft Foundry |
| --- | --- | --- | --- | --- |
| `opus` | Opus 5.5 (from v2.1.280) | Opus 5.5 | Opus 5.5 | Opus 4.6 |
| `sonnet` | Sonnet 5.5 (from v2.1.284) | Sonnet 4.6 | Sonnet 4.5 | Sonnet 4.5 |
| `haiku` | Haiku 5.5 (from v2.1.293, `latest` channel) | Haiku 4.5 | Haiku 4.5 | Haiku 4.5 |

The `fable` alias is Fable 5.1, and `opusplan` uses the `opus` alias in plan mode and the `sonnet` alias for execution, so on the Anthropic API it now pairs Opus 5.5 with Sonnet 5.5. Ultracode, the automatic workflow orchestration, changed in v2.1.284: it is a Claude Code setting, not an effort level. Turning it on or off with `/effort` or the `ultracode` setting leaves the effort level unchanged, while `claude --effort ultracode` turns it on and also sets `xhigh`. Before v2.1.284, turning it on set `xhigh` and an effort cap below `xhigh` made it unavailable; now it is unavailable only when workflows are turned off or the model does not support `xhigh`. To pin models for a team, use the `availableModels`, `enforceAvailableModels` and `maxEffortLevel` settings in managed settings.

**Codex**

GPT-6.1 Sol has been the bundled default since CLI 0.159.1 (2026-09-29), with default effort `low` in Codex; the API default is `medium`. GPT-6 Astra, priority 2 in the catalog, is the pick by hand for the hardest work. The server can override the default for signed-in accounts, so pin the model if you need a fixed one. GPT-6 Sol and Luna default to `medium`. Codex adds an `ultra` level ("Maximum reasoning with automatic task delegation") on GPT-6.1 Sol and Astra; Luna has no `ultra`. OpenAI's guidance: for GPT-6.1 Sol, "start with the reasoning effort available by default in your client", and "Most tasks do not need Max or Ultra."

```bash
# Terminal: one session on GPT-6.1 Sol at high effort
codex -m gpt-6.1-sol -c model_reasoning_effort="high"

# List the models this Codex build knows about
codex debug models --bundled
```

```toml
# ~/.codex/config.toml: pin the team default
model = "gpt-6.1-sol"
model_reasoning_effort = "medium"
model_context_window = 400000   # raise from the 272K default, up to 872K
```

Above 272K input, requests are billed at a higher tier (OpenAI bills GPT-6.1 Sol at $4 / $15 there, against $2 / $10 below it, and GPT-6 Astra at $20 / $75 against $10 / $50). Raise the window only for sessions that need it, for example with `-c model_context_window=400000` on one run.

**Cursor**

Choose the model in the model picker for each chat or agent. Cursor names no single default model (checked 2026-10-12), and Grok 4.7 defaults to `high` effort, with `xhigh`, `medium` and `low` available. Read each model's current rate in Cursor's settings and on its pricing page; the Cursor tab in the price section above lists the rates checked on 2026-10-12. The effort and eval rules below apply unchanged.

## How do you prove a model change was worth it?

A model switch is a change to your production system, so it gets the same evidence as a code change. You do not judge it by reading the agent's diffs; you judge it by what passes your gates.

1. **Pin today's setup.** Record the exact model ID (`claude-opus-5-5`, `gpt-6.1-sol`), the effort level and the tool version, so an alias that moves later does not change behaviour silently.

2. **Sweep effort before models.** Run the same tasks at the next effort level up. Anthropic's Claude Code docs say "Opus 5.5 at `medium` matches or exceeds Opus 5 at `high`", which is the pattern to expect: effort is the cheaper lever.

3. **Run your own eval set.** Use 20–40 tasks taken from your merged pull requests, each with a test command that decides pass or fail. [Building that eval set](/en/ecosystem/models/benchmarks/) is covered on the benchmarks page.

4. **Compare three numbers.** Accepted-task rate (tests, type check and lint pass, and the change survives review), cost per accepted task, and wall-clock time per task.

5. **Roll out one workflow (loop) at a time, with a rollback.** The tech lead who owns the managed settings signs off, changes the pin for one agent workflow at a time, and reverts the pin if the accepted-task rate drops. The [new-model playbook](/en/ecosystem/models/new-model-playbook/) has the full 48-hour procedure.

**Copy-paste prompt: find stale model IDs in a repository**

```text
Search this repository for hard-coded AI model IDs and aliases: CI workflows,
agent settings (.claude/, .codex/, .cursor/), SDK calls, environment files and docs.
Look especially for claude-opus-5 and claude-fable-5 (exact IDs, not claude-opus-5-5
or claude-fable-5-1), claude-sonnet-5 (exact ID, not claude-sonnet-5-5),
claude-haiku-4-5, claude-sonnet-4-* (claude-sonnet-4-5 retires on 2026-11-30),
claude-opus-4-*, gpt-5.6-*, gpt-5.5 (retires from Codex on 2026-10-14), gpt-5.4,
gpt-6-astra and gpt-6-sol pinned as a Codex default, gemini-3.1-pro, gemini-3.5-flash
and gemini-3.7-flash. Treat claude-opus-5-5, claude-fable-5-1, claude-sonnet-5-5,
claude-haiku-5-5, gpt-6.1-sol and gpt-6-luna as current.
Do not edit anything. Output a table with: file:line, the ID found, what calls it
(CI job, subagent, app code, docs), and whether it is pinned on purpose or looks stale.
Finish with the three changes you would make first and the test command that would
prove each change is safe.
```

**Copy-paste prompt: effort sweep on one real task**

```text
Create a branch named effort-sweep-LEVEL from MERGED_SHA^1 and switch to it.
Do not read commit MERGED_SHA or any later commit: it contains the accepted solution.
Implement this task: ISSUE_TEXT
Then run `npm test` and `npm run typecheck`. Report pass or fail for each, the list of
files you changed, and one sentence on anything you were unsure about.
Make one attempt only; if a check fails, report it and stop.
```

Replace `LEVEL` with `medium`, `high` or `xhigh`, `MERGED_SHA` with the merge commit of a real pull request, `ISSUE_TEXT` with its original issue text, and the two commands with your own test and type-check commands. Run the prompt in three fresh sessions, one per effort level: `/effort medium`, `high` and `xhigh` in Claude Code, or `-c model_reasoning_effort="medium"` and so on in Codex. Start each session in its own worktree (`claude --worktree` or `codex --worktree`) so the runs cannot see each other's changes. Read the cost of each run in `/usage` in Claude Code; in Codex, `/status` shows estimated credits or cost for eligible workspaces (0.148.0+) and `/usage` shows account usage (0.156.0+); otherwise compare wall-clock time and pass or fail. Then compare the three reports.

## What do the benchmarks say about these models?

On the official Terminal-Bench 4.0 leaderboard (read 2026-10-12), Claude Opus 5.5 (max effort) running in Claude Code scores 64.8% ± 3.1, Claude Sonnet 5.5 (max) in Claude Code 61.8% ± 2.9, and GPT-6 Astra (max) and GPT-6.1 Sol (max) in Codex 58.2% ± 2.8 and 58.2% ± 3.1. Claude Fable 5.1 (max) in Claude Code scores 57.9% ± 3.8. The Opus 5.5 and Sonnet 5.5 intervals overlap, and Astra and Sol tie, so the board does not separate them. The board also shows effort ladders for some models: GPT-6 Astra in Codex scores 50.6% at `low` and 58.2% at `max`, and Fable 5.1 in Claude Code 43.3% at `low` and 57.9% at `max`. The board has no Claude Haiku 5.5 entry on 2026-10-12.

Anthropic's Sonnet 5.5 launch table (2026-09-28) reports 70.6% for Sonnet 5.5 and 66.4% for Opus 5.5 at `xhigh` effort, the same Opus figure as in its Opus 5.5 launch table (2026-09-22). That puts Sonnet 5.5 above Opus 5.5, while on the official board the order is the reverse. Anthropic itself warns: "at these levels of capability we've found that benchmark margins have become a less reliable guide to real-world differences. In our own use, the gap between Opus 5.5 and Claude Fable 5.1 is narrower than these scores suggest", and for Sonnet 5.5: "Opus 5.5 remains clearly stronger at complex, open-ended work requiring sustained judgment."

**Correction (2026-10-12).** The 2026-09-26 version of this page said that neither Claude Opus 5.5 nor GPT-6 Astra had an entry on the official board. That was wrong for Astra, which has had entries since 2026-09-03: the September read covered only the 21 submission files in the GitHub folder, not the 35-row board on the Harbor Hub. Opus 5.5 joined the board on 2026-10-06.

Every score measures a model inside one agent harness on someone else's tasks. Read [how to read coding-agent benchmarks](/en/ecosystem/models/benchmarks/) before you quote one, and let your own eval set make the decision.

## What breaks after a model change?

**Common failures and how to recover**

- **A Claude model is missing from the Claude Code picker.** You are on a version that predates it: Opus 5.5 needs v2.1.280, Sonnet 5.5 needs v2.1.284 and Haiku 5.5 needs v2.1.293, which only the `latest` channel has on 2026-10-17 (`stable` is v2.1.292). Run `claude update`, or install the latest channel with `claude install latest`.
- **Fable 5.1 drains your usage.** Fable can bill to usage credits depending on plan. To keep subagents off Fable, set `model: sonnet` in each subagent's file under `.claude/agents/`, or set `CLAUDE_CODE_SUBAGENT_MODEL`, and check the weekly Fable limit in `/usage`.
- **API calls return 400 after moving to Opus 5.5 or Fable 5.1.** Thinking cannot be disabled on either model, forced tool use returns an error, and a non-default `temperature`, `top_p` or `top_k` returns 400 on Claude 4.7 and later. Remove those parameters and control cost with `effort` instead.
- **API calls fail after moving from Sonnet 5 to Sonnet 5.5.** Anthropic lists five breaking changes ([Sonnet 5.5 overview](https://platform.claude.com/docs/en/models/sonnet-5-5/overview)): turn up-front thinking off with `thinking: {"type": "between_tools"}`, because `disabled` returns 400; forced tool use (`tool_choice` `any` or `tool`) returns an error; thinking blocks are tied to the model and the conversation; `computer_20251124` is not accepted on the Claude API or Google Cloud; and the advisor tool rejects Opus 4.8, Opus 4.7, Sonnet 5 and Haiku 5.5 as advisors. Text between tool calls also now comes back in `thinking` blocks.
- **API calls fail after moving from Haiku 4.5 to Haiku 5.5.** Manual extended thinking returns an error, sampling parameters can return an error, assistant prefill returns an error, `computer_20250124` errors on the Claude API, Google Cloud and Bedrock, structured outputs are not available on Amazon Bedrock, and changing earlier turns invalidates thinking blocks ([what's new in Haiku 5.5](https://platform.claude.com/docs/en/models/haiku-5-5/whats-new-haiku-5-5)). The same text also counts as about 30% more tokens than on Haiku 4.5.
- **A Claude answer came from a different model.** Requests flagged as cybersecurity fall back to Opus 4.8 and biology-flagged requests to Opus 5 on Opus 5.5 and Fable 5.1. On Sonnet 5.5, cybersecurity-flagged requests re-run on Sonnet 5 and biology-flagged requests end with a refusal, because Sonnet 5.5 has no biology fallback model. Haiku 5.5 has no server-side fallback: its safety classifiers can decline. If a result looks weaker on security work, check which model answered before you blame the prompt.
- **GPT-6.1 Sol calls fail or the model is missing.** In the API, GPT-6.1 Sol takes effort `low`, `medium` (the default), `high`, `xhigh` and `max`, not `none` or `minimal`, and tool calling needs the Responses API: Chat Completions is supported without tool calling. In Codex it is available on Plus, Pro, Business, Enterprise and Edu, off on Enterprise and Edu until an admin enables it; "Free and Go are not included at launch" (OpenAI's Codex models page, checked 2026-10-12).
- **Codex picked a model you did not choose.** The server can override the bundled default for signed-in accounts. Set `model` in `config.toml` or pass `-m`.
- **Codex runs out of context on a large repository.** The default window is 272K. Raise `model_context_window`, up to 872K for the GPT-6 and GPT-5.6 models, for that session only: input above 272K is billed at a higher tier.
- **A pinned model disappears.** GPT-5.4 was retired in Codex on 2026-08-31 and now migrates to GPT-6 Sol. GPT-5.5 retires from ChatGPT, ChatGPT Work and Codex on 2026-10-14 on all plans. Anthropic retires Claude Sonnet 4.5 on 2026-11-30 on Anthropic-operated platforms, and commits to retiring Claude Opus 4.5 no sooner than 2026-11-24 and Haiku 4.5 no sooner than 2026-10-15, with at least 60 days' notice. Re-run the stale-ID prompt above every month.

## Where to go next with model choice

- [Reading coding-agent benchmarks, and building your own evals](/en/ecosystem/models/benchmarks/): the eval set that makes the decision.
- [When a new model ships: a 48-hour evaluation playbook](/en/ecosystem/models/new-model-playbook/): the rollout and rollback procedure.
- [Open-weight and self-hosted models for coding agents](/en/ecosystem/models/open-weight/): GLM, Qwen, Kimi, DeepSeek and Mistral.
- [Model routing by evidence, not brand](/en/developer-scorecard-guide/model-routing/): routing different jobs to different models.
- [Where the model runs: gateways, cloud routes, residency and zero retention](/en/org/model-hosting/): for CTOs choosing a hosting route.
- [Claude Code cost control](/en/claude-code/operate/cost-control/) and [Codex cost management](/en/codex/operate/cost-management/): keeping spend inside budget.

## Sources checked on 2026-10-12

- Anthropic: [models overview](https://platform.claude.com/docs/en/about-claude/models/overview), [pricing](https://platform.claude.com/docs/en/about-claude/pricing), [model deprecations](https://platform.claude.com/docs/en/about-claude/model-deprecations), [Claude Sonnet 5.5 overview](https://platform.claude.com/docs/en/models/sonnet-5-5/overview) and [announcement](https://www.anthropic.com/claude-sonnet-5-5), [Claude Haiku 5.5 overview](https://platform.claude.com/docs/en/models/haiku-5-5/overview), [what's new](https://platform.claude.com/docs/en/models/haiku-5-5/whats-new-haiku-5-5) and [announcement](https://www.anthropic.com/claude-haiku-5-5), the [Claude Opus 5.5 announcement](https://www.anthropic.com/claude-opus-5-5), [Claude Code model configuration](https://code.claude.com/docs/en/model-config), and the npm [`@anthropic-ai/claude-code` dist-tags](https://registry.npmjs.org/@anthropic-ai/claude-code). [Claude Code costs](https://code.claude.com/docs/en/costs) was last read on 2026-09-26.
- OpenAI: [API pricing](https://developers.openai.com/api/docs/pricing.md), the model pages for [GPT-6.1 Sol](https://developers.openai.com/api/docs/models/gpt-6.1-sol.md) and [GPT-6 Astra](https://developers.openai.com/api/docs/models/gpt-6-astra.md), the [Codex models page](https://learn.chatgpt.com/docs/models.md), [Codex release 0.159.1](https://github.com/openai/codex/releases/tag/rust-v0.159.1), the [Codex 0.162.1 bundled catalog](https://github.com/openai/codex/blob/rust-v0.162.1/codex-rs/models-manager/models.json), and the [OpenAI Python SDK model list](https://github.com/openai/openai-python/blob/main/src/openai/types/shared/chat_model.py).
- Google: [latest Gemini models](https://ai.google.dev/gemini-api/docs/latest-model), [Gemini 3.8 Flash](https://ai.google.dev/gemini-api/docs/models/gemini-3.8-flash), [Gemini 3.1 Pro Preview](https://ai.google.dev/gemini-api/docs/models/gemini-3.1-pro-preview), the [Gemini API changelog](https://ai.google.dev/gemini-api/docs/changelog), [Agent Platform pricing](https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing) and the [Gemini 4 Argon announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/).
- xAI: [Grok 4.7](https://docs.x.ai/developers/grok-4-7.md), [release notes](https://docs.x.ai/developers/release-notes.md) and [models](https://docs.x.ai/developers/models).
- Meta and Mistral: [Muse Wonder](https://dev.meta.ai/models/muse-wonder) and [Mistral Large 4](https://docs.mistral.ai/models/mistral-large-4-0).
- Cursor: [models and pricing](https://cursor.com/docs/models-and-pricing), [available models](https://cursor.com/help/models-and-usage/available-models) and [Grok 4.7](https://cursor.com/docs/models/grok-4-7).
- Benchmarks: the [official Terminal-Bench 4.0 leaderboard](https://www.tbench.ai/?version=4) (Harbor Hub data embedded in tbench.ai).
