Claude Opus 5.5 changelog

Perplexity Agent API xhigh effort raises max steps from 100 to 300

The xhigh level stays on Claude Opus 5.5 with GPT-6.1 Sol as fallback. Context compaction starts at 150,000 tokens, and it no longer sends a system prompt.

Who is affected:

Agent API users on xhigh effort or frozen configurations.

What to do:

Copy the updated values if you use a frozen configuration.

Full Claude Platform available on Google Cloud (Anthropic managed)

Claude Models on Google Cloud (Anthropic managed) is generally available with US and Global endpoints. It gives Google Cloud customers the full Claude Platform feature set with Google Cloud IAM, billing and commitment drawdown. Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5 are available, and fast mode for Opus 5.5 costs $8 / $40 per million input and output tokens there.

Who is affected:

Teams that buy Claude through Google Cloud and want features that the Google-managed offering lacks.

What to do:

No action needed. Google Cloud customers who need same-day Claude API features can switch to the Anthropic-managed offering.

Vals Index: Sonnet 5.5 edges Opus 5.5 and Gemini 4 Argon

Vals AI's GDP-weighted index of finance, coding, legal and tax tasks ranks Claude Sonnet 5.5 first at 70.85%, Opus 5.5 at 70.18%, Gemini 4 Argon 69.24%, GPT-6 Astra 65.02% and GPT-6.1 Sol 63.94% as of October 21, 2026.

Arena text leaderboard: Gemini 4 Argon first, Opus 5.5 second, GPT-6 Astra 73rd

Arena's October 20, 2026 text leaderboard ranks Gemini 4 Argon (High) first at 1532 (4,900 votes), Claude Opus 5.5 (High) second at 1516, Claude Opus 5 (Max) fourth at 1509, Claude Sonnet 5.5 (xHigh) 28th at 1473, GPT-6.1 Sol (Max) 66th at 1443 and GPT-6 Astra (Max) 73rd at 1441.

Livenerf finds no change in Opus 5.5 accuracy in its first 10-day window

Window 1 (days 11 to 20, from October 4) scored 0.592 against a 0.584 launch-week baseline, a difference of +0.007 plus or minus 0.018 over 780 samples, which the pre-registered rule calls no change. Median output rose from 588 to 606 tokens. Day 19 ran on API credits because the Max weekly meter was exhausted.

Arena WebDev leaderboard: Opus 5.5 first, GPT-6 Astra second

Arena's October 19, 2026 WebDev leaderboard ranks Claude Opus 5.5 (Max) first at 1807 (2,925 votes), GPT-6 Astra (Max) second at 1774, Claude Sonnet 5.5 (xHigh) third at 1768, GPT-6.1 Sol (Max) fourth at 1753, Claude Opus 5 (Max) eighth at 1689 and Gemini 4 Argon (High) tenth at 1678.

Vals AI Terminal-Bench 4.0 with one harness: Opus 5.5 65.2%, Astra 59.6%, Argon 57.6%

Running every model in the same Mini-SWE-agent harness, Vals AI reports Claude Opus 5.5 at 65.15%, Sonnet 5.5 64.14%, GPT-6 Astra 59.60%, Gemini 4 Argon 57.58%, GPT-6.1 Sol 55.05% and Opus 5 53.53%. GPT-6.1 Sol cost $1.72 per test against $13.20 for Opus 5.5.

Fast mode for Claude Opus 5.5 leaves research preview

Fast mode on Opus 5.5 no longer needs the beta header or an access request, gets a default rate limit in the Start, Build and Scale tiers, and is available on Microsoft Foundry. Fast mode costs $8 / $40 per million tokens.

Antigravity 2.24 adds model choice for automations and Claude 5.5 models through Vertex AI

Antigravity app 2.24.0 lets automations run on Flash Lite, Flash or Pro, adds /compact and a redesigned side panel. With Application Default Credentials sign-in, the model picker offers Claude Opus 5.5 and Claude Sonnet 5.5 when the Google Cloud project has access to them.

SlopCodeBench: GPT-6 Astra leads strict pass rate, Opus 5.5 does better at medium than xhigh

On 36 problems and 196 checkpoints, GPT-6 Astra at xhigh had the best strict pass rate (16.33%), GPT-6.1 Sol medium 14.80% and Opus 5.5 medium 14.29%. Opus 5.5 at xhigh scored lower (12.76%) at 4.22 times the cost. The Hacker News thread drew 303 points and 120 comments.

Who is affected:

Teams choosing effort levels for long coding tasks with changing requirements.

What to do:

Test medium effort before paying for xhigh.

Opus 5.5 in Claude Code tops the official Terminal-Bench 4.0 leaderboard at 64.8%

Claude Code with Opus 5.5 at max effort resolved 64.85% of Terminal-Bench 4.0 tasks (plus or minus 3.1, 330 trials, $4,718.65 total cost), ranked first. The entry was posted October 6, 2026.

GitHub issue claims Opus 5.5 quality fell back to Opus 5 level

A single Claude Code GitHub issue on version 2.1.284 claims Opus 5.5 quality degraded to the level of Opus 5. It was labeled needs-info and has no Hacker News thread. Livenerf found no change in the same period.

Claude Opus 5.5 now powers Standard effort in Perplexity Computer

Claude Opus 5.5 became the model behind Standard effort in Computer for eligible paid personal accounts. GPT-6.1 Sol now powers Light effort.

Epoch: Opus 5.5 takes the top spot on the Epoch Capabilities Index

Epoch AI reported Opus 5.5 first on its Epoch Capabilities Index with 167, narrowly ahead of GPT-6 Astra, and Sonnet 5.5 roughly matching Fable 5.1 (165). Epoch's current ECI data lists Opus 5.5 at 168.83, GPT-6 Astra 168.81, GPT-6.1 Sol 168.37, Gemini 4 Argon 166.45, Sonnet 5.5 166.18 and Opus 5 163.94.

Livenerf starts a 30-day test of whether Opus 5.5 gets worse after launch

Livenerf runs a frozen 78-question panel (from GPQA Diamond, MMLU-Pro, competition math and AIME) through headless Claude Code on a Max plan, 90 samples a day for 30 days starting September 24, 2026, with a pinned CLI and exact-match grading. The Hacker News launch thread drew 926 points and 394 comments.

Who is affected:

Claude Max subscribers who suspect Opus 5.5 has been quietly downgraded.

What to do:

No action needed. Watch for the first decision around October 24, 2026.

MarginLab Claude Code tracker collects an Opus 5.5 baseline, with detection paused

As of its September 29, 2026 update, MarginLab's Claude Code tracker (Opus 5.5 at high effort, from September 24) showed a 77% daily pass rate on 50 tasks, 76% over 7 days (350 tasks) and 81% over 30 days (1,200 tasks). Degradation detection is paused until a baseline exists.

Who is affected:

People watching Claude Code with Opus 5.5 for silent regressions.

What to do:

No action needed. Check the tracker again once the baseline is set.

Claude Code defaults to Opus 5.5, including on Pro plans

Claude Code 2.1.280 made Opus 5.5 the default Opus model and changed the default on Pro and Team Standard plans from Sonnet to Opus.

Who is affected:

Claude Code users on Pro and Team Standard plans, whose usage limits may now be consumed by Opus.

What to do:

Use /model to pick Sonnet if you want to stretch usage limits.

Claude Opus 5.5 released at a lower price than Opus 5

Claude Opus 5.5 launched at $4 / $20 per million tokens (Opus 5 is $5 / $25), with cache reads at $0.20 per million tokens and always-on adaptive thinking.

Who is affected:

Developers and subscribers using Claude Opus 5.

What to do:

Test claude-opus-5-5. Remove any thinking disabled setting and forced tool_choice first.

Five-hour usage limits raised with Opus 5.5, plus a saved limit reset

Alongside Opus 5.5, Anthropic raised five-hour usage limits on Pro, Max, Team and seat-based Enterprise plans and gave subscribers one rate limit reset they can use when they choose.

Claude Opus 5.5 cannot turn thinking off

On Opus 5.5, thinking disabled and manual thinking both return a 400 error, forced tool_choice returns a 400 error, and the older computer use tool is rejected on the Claude API and Google Cloud.

Who is affected:

Code moving from Claude Opus 5 to Opus 5.5.

Action needed:

Omit the thinking field, control depth with effort, and use tool_choice auto.

Claude Opus 5.5 available in GitHub Copilot

Claude Opus 5.5 is available on Copilot Pro+, Max, Business and Enterprise at provider list price. GitHub says it resolved tasks about as well as Opus 5 with significantly fewer steps and tokens.

New API accounts can no longer edit context around thinking blocks

To counter distillation, the API now checks that a thinking block is sent back with the same system prompt, tools and messages. On Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5, accounts created on or after August 31, 2026 get an error unless they opt into dropping the block.

Who is affected:

Agents that compact or rewrite earlier turns, especially on new API accounts.

Action needed:

Use server-side compaction or mid-conversation tool changes, or opt into non-strict mode.