AnthropicPrevious version · Claude Opus 5.5
Claude Opus 5
Anthropic pitches Opus 5 as "close to the frontier intelligence of Claude Fable 5 at half the price" — but that price is just Claude Opus 4.8's unchanged $5/$25 rate carried over, not a cut. What's real is the scorecard: fourteen of the sixteen benchmark results tracked here are independent runs — the exceptions are OSWorld 2.0 and Toolathlon-Verified figures from Anthropic's own system cards, not independent runs.
Claude Opus 5 benchmarks and pricing, every number sourced: 16 tracked Claude Opus 5 benchmark scores (14 independently run, 2 still resting on a vendor’s own claim), priced at $5.00 per million input tokens and $25.00 per million output.
- Released
- 2026-07-24
- License
- proprietary
- Context window
- 1M tokens
- Knowledge cutoff
- 2026-05
- Verified
- sources checked 2026-08-17–2026-10-25
- Succeeded
- Claude Opus 4.8 (2026-05-28)
- Parameters
- Not disclosed
- Architecture
- Not disclosed
Claude Opus 5’s verified record
Claude Opus 5’s featured comparison is Claude Sonnet 5: 3 leads, 1 tie, and 7 not callable. Claude Opus 5 is priced at $5.00/$25.00 per 1M tokens (in/out) vs Claude Sonnet 5’s $2.00/$10.00. Full verdict →
Against the 324 head-to-head comparisons Claude Opus 5 shares with other tracked models:
A gap counts for Claude Opus 5 only where both sides were run independently and the benchmark still separates models. Losses are listed alongside wins on purpose — Claude Opus 5 trails on 5 of them.
HLE · no tools Reasoning
±2 is noiseLiveBench Composite score across 7 domains
±2.7 is noiseAnalystAgent Spreadsheet & document analysis
±11.2 is noiseTerminal-Bench 4.0 · max Terminal ops
±12.4 is noiseDeepSWE Long-horizon coding
±9.5 is noiseARC-AGI-2 · max Compositional visual reasoning
±9.2 is noiseAgents' Last Exam Professional work
±3.2 is noiseNo verdict for Claude Opus 5 anywhere on Terminal-Bench 2.1, OSWorld 2.0 · v2026 08 08 full, OSWorld 2.0 · v2026 08 08 offline ( every independently confirmed comparison inside the noise band); GPQA Diamond, LiveCodeBench, SWE-bench Verified ( saturated); OSWorld 2.0 · v2 1 full, OSWorld 2.0 · v2026 08 08 full modified, Toolathlon-Verified ( nothing independently confirmed on both sides).
What changed from Claude Opus 4.8 to Claude Opus 5
The 11 benchmarks both models have been scored on, using the same variant each time. A raw Claude Opus 5 gain is not a real gain until it clears that benchmark’s own noise band, so each row below carries the verdict and not just the arithmetic.
| Benchmark | Claude Opus 4.8 | Claude Opus 5 | Change | Verdict |
|---|---|---|---|---|
| HLE | 48.7 | 54.9 | +6.2 | Real gap |
| Terminal-Bench 2.1 | 78.9 | 89.1 | +10.2 | Setup-dependent |
| DeepSWE | 59 | 74 | +15.0 | Real gap |
| Toolathlon-Verified | 76.2 | 80.6 | +4.4 | Unverified |
| Agents' Last Exam | 27 | 32.2 | +5.2 | Setup-dependent |
| SWE-bench Verified | 88.6 | 97 | +8.4 | Tainted |
| GPQA Diamond | 92.42 | 93.43 | +1.0 | Tainted |
| LiveCodeBench | 87.82 | 89.03 | +1.2 | Tainted |
| LiveBench | 76.2 | 80.1 | +3.9 | Real gap |
| AnalystAgent | 45 | 53.75 | +8.8 | Tie |
| Terminal-Bench 4.0 | 23.23 | 53.54 | +30.3 | Real gap |
Claude Opus 5 API pricing
$5.00 in / $25.00 out per 1M tokens — official pricing source
What Claude Opus 5 costs per job
| Workload | Tokens in / out | Cost |
|---|---|---|
| One long chat turn | 100K / 10K | $0.750 |
| A codebase review | 1,000K / 100K | $7.50 |
| A day of agent work | 10,000K / 1,000K | $75.00 |
Computed from Claude Opus 5’s list rates above — cache discounts and batch tiers are not applied.
Claude Opus 5 is one of 9 Anthropic models tracked on this site, at these official list prices.
| Model | In / 1M | Out / 1M |
|---|---|---|
| Claude Fable 5.2 | $8.00 | $40.00 |
| Claude Haiku 5.5 | $0.10 | $0.50 |
| Claude Sonnet 5.5 | $2.00 | $10.00 |
| Claude Opus 5.5 | $4.00 | $20.00 |
| Claude Fable 5.1(previous version) | $10.00 | $50.00 |
| Claude Opus 5(previous version) | $5.00 | $25.00 |
| Claude Sonnet 5(previous version) | $2.00 | $10.00 |
| Claude Fable 5(previous version) | $10.00 | $50.00 |
| Claude Opus 4.8(previous version) | $5.00 | $25.00 |
Claude Opus 5 benchmark scores
| Benchmark | Score |
|---|---|
| HLE(no tools)[1] Reasoning · ±2 is noise | |
| Terminal-Bench 2.1[2] Terminal ops · ±10.6 is noise | |
| DeepSWE[3] Long-horizon coding · ±9.5 is noise | |
| SWE-bench Verifiedsaturated[4] Bug fixing — not ranked at any gap size | |
| LiveCodeBenchsaturated[5] Contest coding — not ranked at any gap size | |
| GPQA Diamondsaturated[6] Expert science Q&A — not ranked at any gap size | |
| Agents' Last Exam[7] Professional work · ±3.2 is noise | |
| ARC-AGI-2(max)[8] Compositional visual reasoning · ±9.2 is noise | |
| LiveBench[9] Composite score across 7 domains · ±2.7 is noise | |
| AnalystAgent[10] Spreadsheet & document analysis · ±11.2 is noise | |
| Terminal-Bench 4.0(max)[11] Terminal ops · ±12.4 is noise | |
| OSWorld 2.0(v2 1_full)[12] Long-horizon computer use · ±9.7 is noise | |
| OSWorld 2.0(v2026 08_08_full)[13] Long-horizon computer use · ±9.7 is noise | |
| OSWorld 2.0(v2026 08_08_offline)[14] Long-horizon computer use · ±9.7 is noise | |
| OSWorld 2.0(v2026 08_08_full_modified)[15] Long-horizon computer use · ±9.7 is noise | |
| Toolathlon-Verified[16] Multi-tool chores · ±9.7 is noise |
Who ran these numbers: 14 of 16 independent — vals.ai (4), artificialanalysis.ai (3), osworld-v2.xlang.ai (3), deepswe.datacurve.ai (1), snorkel.ai (1), arcprize.org (1), livebench.ai (1); vendor self-reported (2).
- HLE: AA live leaderboard value, 'Claude Opus 5 (Adaptive Reasoning, Max Effort)' — text-only, no tools. Re-verified on the board 2026-08-19. Corrected 2026-08-19: this row used to cite AA's Opus 5 launch article, which reports 53.0 — the 54.9 was always the board's number, so the citation pointed at a page that did not support it. Anthropic's own max-effort run is 56.3 (card §8.1.A) — not stored as a row; this cell holds AA's independent run.
- Terminal-Bench 2.1: AA's own run, chart-JSON label 'Claude Opus 5 (max)'; page prose gives the fuller label 'Claude Opus 5 (Adaptive Reasoning, Max Effort)' for the same score. A lower-effort 'Claude Opus 5 (xhigh)' variant also appears in the same chart at 88.0. Not on tbench.ai's official board, which lists only older Anthropic models (Fable 5, Opus 4.8, Opus 4.7, Sonnet 5). Cross-check (2026-10-01): vals.ai's archived Terminus 2 table lists it at 84.64, with Opus 4.8 as a refusal fallback on 9 of 267 tasks; counting those as failures gives 81.27.
- DeepSWE: 74%±4 pass rate on the mini-swe-agent harness (all models run identically); board updated 2026-08-13. Anthropic's card reports 68.8 (max, five trials) — not stored as a row; this cell holds the Datacurve board's independent 74.
- SWE-bench Verified: vals.ai run, rank 1 of 83 (updated 2026-08-14); Anthropic self-reports 96.0. Benchmark near-saturated — five models at 95%+. Anthropic's card reports 96.0 (max, five trials) — not stored as a row; this cell holds vals.ai's 97.
- LiveCodeBench: vals.ai, updated 2026-08-15; re-verified on the vals.ai model page 2026-08-17.
- GPQA Diamond: vals.ai run, rank 7/135 (board updated 2026-08-19).
- Agents' Last Exam: Overall pass rate 32.2 (Claude Code, High, the board's best effort for this model; score 55.9; est. cost $1,108). Per-effort results on the board: High 32.2, Max 30.9, XHigh 30.3, Medium 28.9, Low 27.6. Snorkel's data notes record that the board's upstream source re-published every Claude Opus 5 figure on 2026-09-24, moving its Overall best tier (High) from 31.6 to 32.2. This site's earlier 27.0 (Max; score 49.0; est. cost $7,239, recorded 2026-08-20) no longer appears; the current Max run reads 30.9 (score 52.7).
- ARC-AGI-2: Claude Opus 5's official ARC-AGI-2 leaderboard row, dated 2026-07-24 on arcprize.org. Higher than its only other populated tier, High (88.3%).
- LiveBench: Board row "Claude 5 Opus Thinking Max Effort (LiveBench's own row label reverses this site's word order; corroborated as the same model by a real GitHub issue on livebench/livebench)" on the 2026-06-25 LiveBench release.
- AnalystAgent: AA's own run, board row 'Claude Opus 5 (Adaptive Reasoning, Max Effort)'; the board rounds this to 53.8. pass@1 63.75, pass@5 73.75.
- Terminal-Bench 4.0: vals.ai board row — mini-swe-agent harness (single bash tool), pass@1 averaged over 3 full passes, raw value 53.535 at effort max. Official board (grant-funded, Claude Code, xhigh effort, 5-trial pass@1): 53.94 ± 3.17 CI — 0.4 points above this run. The official board's best tier for this model is xhigh; its max-effort official row is 51.82 and its rows fall to 44.85 (medium) and 34.85 (low).
- OSWorld 2.0: Authors' own run on the v2.1 release, full environment. Partial reward 77.67, binary accuracy 44.33; batch tool calls, 500 steps. Effort ladder on the same slice: 59.09 / 64.85 / 68.92 / 73.11 / 77.67 partial (low to max). Anthropic's own-harness rerun of the same model on its September 10 task files reports 74.0 partial / 37.2 strict — three-plus points under this row under a different protocol, so the two are not comparable (see the v2026_08_08_full_modified row for that pattern). Verified twice on 2026-10-08.
- OSWorld 2.0: Authors' own run on the August 2026 release, full environment. Partial reward 68.31, binary accuracy 31.43; batch tool calls, 500 steps. Effort ladder: 54.89 / 60.49 / 63.79 / 67.67 / 68.31 partial (low to max). This is the slice directly comparable with the GPT-5.6 Sol full-set row below. Verified twice on 2026-10-08.
- OSWorld 2.0: Authors' own run, August release, offline subset (the tasks that work without internet access). Partial reward 70.19, binary accuracy 34.72. This is the row OpenAI's GPT-6 Astra page points to: its footnote says Claude numbers are the authors' leaderboard reproduction, and it quotes Opus 5 at 70.2 — a match to a decimal. Effort ladder on this slice: 55.22 / 60.26 / 65.92 / 70.14 / 70.19 partial (low to max). Verified twice on 2026-10-08.
- OSWorld 2.0: Anthropic's rerun under its own protocol (Fable/Mythos 5.1 launch materials, 2026-09-01, detailed in that system card §8.14.3): the authors' August 2026 task release plus Anthropic's own task-setup and grading-script fixes, Claude Opus 4.8 substituted as model grader, 1080p, 500 steps, pass@1 over five runs. Strict pass rate 39.6. Compare with the same model's official August full-set row (68.31): a 7.1-point difference attributable to the modified protocol, which is why this slice has its own variant. OpenAI and Google both declined to carry these figures. Verified on the Opus 5 launch page grid and in Opus 5's system card text.
- Toolathlon-Verified: System card Table 8.13.6.A, 80.6% Pass@1 at max effort against the Verified release, explicitly with no safety classifiers or fallback (unlike the later models' production-config runs).
Notes on the record
Successor to Claude Opus 4.8 at the same $5/$25 price (per Anthropic's pricing and model-overview pages, checked 2026-08-20). A previous version since Claude Opus 5.5 shipped on 2026-09-22, and cheaper: this site's rule is same product line and price tier, newer release, and Opus 5.5 lists $10% below Opus 5's $5/$25. Anthropic's own model page agrees from its side, calling Opus 5 "a legacy model" and naming 5.5 the current Opus (checked 2026-10-04). What Anthropic has not done is set a retirement date — its deprecations page still lists Opus 5 "Active", a status that page's own glossary defines as "fully supported and recommended for use", with a tentative retirement no earlier than 2027-07-24. Configurable 'effort' toggle (shared machinery, first shipped on Claude Opus 4.7) — Anthropic's docs list five levels (low, medium, high, xhigh, max), defaulting to high on the Claude API and Claude Code (source: platform.claude.com/docs/en/about-claude/models/overview). Anthropic positions it as 'close to the frontier intelligence of Claude Fable 5 at half the price'.
Max output 128k (300k via batch beta, using the `output-300k-2026-03-24` beta header on the Message Batches API, per the same source). Also available via Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry.
The card's OSWorld 2.0 figure of 70.6 (500-step budget, old task set) is not version-mappable — the official cards state the old 2.0 set and the v2.1 set are not comparable.
Compare with
FAQ
Has Claude Opus 5 been independently benchmarked?
Yes — on fourteen of the sixteen scores tracked here — the exceptions are Anthropic's own claimed OSWorld and Toolathlon numbers. The fourteen independent results — Humanity's Last Exam (no tools), Terminal-Bench 2.1, DeepSWE, SWE-bench Verified, LiveCodeBench, GPQA Diamond, Agents' Last Exam, ARC-AGI-2, LiveBench, AA-AnalystAgent, and Terminal-Bench 4.0 (vals.ai, added 2026-10-08) — are independent runs, not Anthropic's own reported figures: HLE, Terminal-Bench 2.1 and AA-AnalystAgent are from Artificial Analysis (observed 2026-08-19, 2026-08-20 and 2026-09-29), DeepSWE is from DataCurve (observed 2026-08-17), Agents' Last Exam is from Snorkel AI (re-read 2026-10-01), SWE-bench Verified, LiveCodeBench, and GPQA Diamond are all from Vals.ai (observed 2026-08-17 and 2026-08-20), and ARC-AGI-2 and LiveBench are from ARC Prize and LiveBench's own board (both observed 2026-08-24).
Why is Claude Opus 5 marked as a previous version?
Because Claude Opus 5.5 shipped on 2026-09-22, in the same line and price tier — which is how this site decides, not what a vendor says. Anthropic's own page backs it from its side too, calling Opus 5 a legacy model and naming 5.5 the current Opus. What Anthropic has not done is set a retirement date: its deprecations page still lists Opus 5 "Active", which that page's glossary defines as fully supported and recommended for use, with a tentative retirement no earlier than 2027-07-24. Opus 5 therefore stays on sale at $5/$25 and keeps every score on this page — previous version means newer in the lineup, not switched off.
Does 'half the price' mean Claude Opus 5 is cheaper than Claude Opus 4.8?
No. Claude Opus 5 is priced at $5 per million input tokens and $25 per million output tokens — identical to Claude Opus 4.8's rate, per Anthropic's pricing docs. The 'half the price' language in Anthropic's launch messaging compares Opus 5 to Claude Fable 5, priced at $10/$50 per million tokens on the same Anthropic models-overview page — a separate and pricier model in the current lineup, not a discount off its own Opus predecessor.
What is the effort toggle on Claude Opus 5, and which level is the default?
Claude Opus 5 carries the configurable 'effort' parameter — low, medium, high, xhigh, or max — first introduced with the full five-level scale on Claude Opus 4.7 in April 2026. It trades reasoning depth and token spend against speed and cost. One thing Opus 5 does change: at xhigh or max, thinking can no longer be disabled at all, a breaking change from Opus 4.8. Anthropic's documentation sets the default to high on the Claude API and in Claude Code; that's narrower than Claude Opus 4.8, where high is the default across every surface including claude.ai, so effort may need to be set explicitly elsewhere for Opus 5.
Can Claude Opus 5 generate more than 128,000 output tokens in one response?
Not through the standard Messages API, which caps Claude Opus 5 at 128K output tokens. Anthropic's Message Batches API supports up to 300K output tokens for Opus 5 using the `output-300k-2026-03-24` beta header — the batch-only extended-output mode this model's notes refer to.
Does Claude Opus 5 support the same extended-thinking controls as older Claude models?
No. Claude Opus 5 replaces the older manual 'extended thinking' toggle (with a fixed token budget) with adaptive thinking, tuned instead via the effort parameter. Anthropic's own model comparison table lists extended thinking (`thinking.type: "enabled"`) as unsupported on Opus 5, with adaptive thinking supported in its place.
Further reading
- GPQA Diamond leaderboard 2026 — Claude Opus 5 is one of the 23 models it compares.
- Models with 10M token context windows 2026 — Claude Opus 5 is one of the 51 models it compares.
Benchmark guides
- Humanity's Last Exam — what Claude Opus 5’s reasoning score on it does and doesn’t prove.