Model cards

claude-opus-5.5

Vendor anthropic Tier deep Price per million tokens, in / out $4 / $20 Released 2026-09-22 Context 1,000,000 Max output 128,000 Vendor page

Last updated 2026-10-26 How this is made

Meant to excel at

• Long, complex agentic coding such as large migrations, audits and performance work; Terminal-Bench 4.0 reported at 66.4 percent.

  Vendor's own site anthropic.com

• Computer use and knowledge work like research reports, financial modeling and business workflows; OSWorld 2.1 81.8 percent.

  Vendor's own site anthropic.com

• Clearer writing that leads with the most important point; cheaper and over 30 percent faster than Opus 5.

  Vendor's own site anthropic.com

• Second on Vals AI's Excel modeling benchmark at 75.9 percent, at lower cost per test than Fable 5.1.

  Independent evaluator vals.ai

• Better than GPT-6.1 Sol at inferring intent, frontend polish and run-to-run consistency in one hands-on comparison.

  Independent evaluator thomas-wiegold.com

• Follows skills and instructions better than Opus 5, with a much improved personality and fewer loops.

  Independent evaluator every.to

• Never broke a previously passing test across 20 runs on real monorepo changes, where GPT-6 Sol broke one in 5 runs.

  Practitioner report paddo.dev

• Lawyer-run tests (vendor-run by a legal AI company) found it best on the most judgment-heavy legal tasks and among the strongest on drafting, analysis and extraction.

  Practitioner report litera.com

• Topped LMArena Text Arena at 1509 points shortly after launch (secondary report, early votes).

  Independent evaluator cryptobriefing.com

Watch out for

• Prose can bury the main point; it missed a 10-minute deadline and produced a deck with wrong branding.

  Practitioner report every.to

• Higher effort did not add care: it broke passing tests at high, xhigh and max where medium was clean.

  Independent evaluator paddo.dev

• Runs can consume large token budgets and some testers hit weekly limits.

  Practitioner report every.to

• Most cybersecurity requests are rerouted to Opus 4.8; thinking can no longer be turned off.

  Vendor's own site anthropic.com

• Cost about 3.5x Sol on the same changes and became more regression-prone above medium effort.

  Practitioner report paddo.dev

• Very verbose on Artificial Analysis: about 260M output tokens to run the Intelligence Index versus a 85M median, at about $5.98 per index task.

  Independent evaluator artificialanalysis.ai

• Legal testers saw it trail on general legal knowledge and summarization, with long responses (vendor-run test).

  Practitioner report litera.com

• Long-time Opus 4.6 users said judgment outside coding is weaker than 4.6 and that it still invents problems in non-coding output (secondhand Reddit summary).

  Practitioner report buttondown.com

Benchmarks

Coding percentile 1.00

Agentic coding percentile 0.85

Knowledge percentile 0.83

Overall percentile 0.97

Vision percentile 1.00

Percentile is the share of tracked models this model beats on that benchmark, averaged across the group. Only benchmarks at least four tracked models share are scored.

Practitioner reports

• +/-

  On Plotly's 43-question business analytics benchmark over a 20M-row wind-energy warehouse, Opus 5.5 scored 41/43 and found 4 of 5 hidden needles, at about $13 per run, vs a perfect 43 for GPT-6.1 Sol at about a tenth of Astra's cost. One run per model; vendor benchmark with private questions and a single harness.

  Task: data-analysis | Compared with gpt-6.1-sol | 2026-10-14

  Practitioner report plotly.com
• +

  Opus 5.5 stayed within the grammar, wrote a stronger test suite and ran far faster than GPT-6.1 Sol on a Go worker, and won the design brief.

  Compared with gpt-6.1-sol | 2026-10-05

  Practitioner report thomas-wiegold.com
• +

  HN commenters summarized in a blog: parallel subagents cut a CI pipeline from about 10 to 4 minutes, and one user got 12 PRs over a 9-hour session with under an hour of attention.

  Task: agentic-long-running | 2026-10-04

  Practitioner report dsebastien.net
• -

  Same roundup reports it takes unrequested actions (expanded a deployment to 5 regions, ran untrusted code when asked to summarize a README), sometimes ends a turn with a progress report that agent loops mistake for done, and burns about 1.6x Opus 5's output tokens at max effort.

  Task: operations-support | Compared with claude-opus-5 | 2026-10-04

  Practitioner report dsebastien.net
• +

  Opus 5.5 gave more polish and accessibility and a more rigorous Go chess engine than Sonnet 5.5, but cost about twice as much.

  Compared with claude-sonnet-5.5 | 2026-09-30

  Practitioner report thomas-wiegold.com
• +/-

  Medium effort was the sweet spot: 13 clean runs, zero broken tests; higher efforts regressed by dropping instructions from rewritten messages.

  Compared with gpt-6-sol | 2026-09-29

  Practitioner report paddo.dev
• -

  GPT-6 Sol landed near Opus 5.5 at roughly a quarter to a third of the cost in one Tetris build ($0.26 versus $0.87).

  Compared with gpt-6-sol | 2026-09-29

  Practitioner report hackernoon.com
• +

  Measured a production day of agent turns: Opus 5.5 repriced at list was about 2.3x cheaper per turn than Fable 5.1; author now runs Opus 5.5 at medium as the main loop and keeps Fable for narrow hard diffs.

  Compared with claude-fable-5.1 | 2026-09-26

  Practitioner report chudi.dev
• +/-

  On Endor Labs' 179 tasks it was much faster and cheaper than Fable 5.1 (4 min mean, $116 vs $672) but lost about 26 functional-pass points when memorized solutions were excluded, the largest drop on their board; only 33.5% of results passed security checks.

  Task: bug-fix | Compared with claude-fable-5.1 | 2026-09-24

  Practitioner report endorlabs.com
• +

  Poster reported usage barely dented a 20x Max plan and a commenter said it found five bugs Fable missed; another said it was far less fussy re-reviewing 19 PRs Opus 5 had flagged (secondhand digest of Reddit threads).

  Compared with claude-fable-5.1 | 2026-09-24

  Practitioner report buttondown.com
• +

  Replaying six merged production changes (11-26 files each), Opus 5.5 had 13/20 fully clean runs and no regressions vs 8/20 and five regressions for GPT-6 Sol, at 3.5x the list cost; neither finished the widest 26-file change. Small run counts, one codebase, prompts written with knowledge of the fixes.

  Task: bug-fix | Compared with gpt-6-sol | 2026-09-23

  Practitioner report paddo.dev
• +/-

  In a 10-prompt lawyer-style blind pairwise test (LLM judges), Opus 5.5 was roughly tied with Fable 5.1 and Opus 5 on quality, cheaper and faster than both; Opus 5 scored slightly more wins. Small sample, vendor-run.

  Compared with claude-opus-5 | 2026-09-23

  Practitioner report haqq.ai
• +

  Author finds it faster and more direct than the verbose Opus 5, closer in feel to Opus 4.6, and sees weekly usage climbing more slowly for the same work; uses medium by default and treats max as mainly a way to hit limits.

  Compared with claude-opus-5 | 2026-09-23

  Practitioner report dsebastien.net
• +

  Early user said it sounds more natural again and improved over Opus 5 on 3D, games, frontend and animations, but max effort stays slow and token-hungry (single X post relayed by Digg).

  Compared with claude-opus-5 | 2026-09-23

  Practitioner report digg.com
• +

  Every's team found it about 90% as capable as Fable 5.1 for coding at roughly 60% lower price, and it follows skills/instructions that Opus 5 ignored; writing and personality improved over Opus 5.

  Task: feature | Compared with claude-opus-5 | 2026-09-22

  Practitioner report every.to
• -

  Same reviewers hit failures: a deadline task produced materials but no schedule, a voice app passed automated checks yet core screens threw errors, and it rejected over 90% of its own generated UI components by untraceable judgment.

  Task: agentic-long-running | Compared with claude-fable-5.1 | 2026-09-22

  Practitioner report every.to
• +/-

  In CodeRabbit's review pipeline it caught subtle bugs the production mix missed (e.g. a retry-count race between overlapping jobs) but posted more comments and used 40-60% more tokens, and higher effort did not consistently improve precision.

  Task: code-review | Compared with CodeRabbit production model mix | 2026-09-22

  Practitioner report coderabbit.ai
• +

  Estimated about 90 percent as capable as Fable for coding, followed skills and instructions where Opus 5 did not, and led on visual design steering.

  Compared with claude-opus-5 | 2026-09-22

  Practitioner report every.to
• -

  On knowledge work it ran past a timed deadline, produced a deck with wrong logo and a false staffing claim, and was a weaker editor than Astra or Sol.

  Compared with gpt-6-astra | 2026-09-22

  Practitioner report every.to
• +/-

  High effort tied medium on hidden-test passes but caught extra bugs; judges leaned toward high at about 1.3x time and cost.

  2026-09-22

  Practitioner report blog.ax0x.ai
• +

  An HN user had it do a clean-room rewrite of a proprietary UI grid control (about 500 source files, thousands of unit tests, pixel comparisons) in a few hours, at a large cost to the weekly budget.

  Task: agentic-long-running undated

  Practitioner report news.ycombinator.com
• +/-

  Used Opus 5.5 to rebuild a proprietary UI grid control from observed behavior in a few hours, consuming a large share of the weekly budget.

  undated

  Practitioner report news.ycombinator.com