Anthropic
Claude Opus 5.5
Anthropic's self-described "new leading model" at $4/$20 per million tokens, 20% below Opus 5. It edges Fable 5.1 on Humanity's Last Exam (61.4% vs 59.1%, just past the noise band — the highest HLE score AA has recorded) and ties it on LiveBench (83.2 vs 83.4). It tops Agents' Last Exam at 38.2 (Max); the Opus 5 row tracked here is a 32.2 High run, so that pairing is setup-dependent.
Claude Opus 5.5 benchmarks and pricing, every number sourced: 10 tracked Claude Opus 5.5 benchmark scores (7 independently run, 3 still resting on a vendor’s own claim), priced at $4.00 per million input tokens and $20.00 per million output.
- Released
- 2026-09-22
- License
- proprietary
- Context window
- 1M tokens
- Knowledge cutoff
- 2026-06
- Verified
- sources checked 2026-09-23–2026-10-25
- Succeeded
- Claude Opus 5 (2026-07-24)
- Parameters
- Not disclosed
- Architecture
- Not disclosed
Claude Opus 5.5’s verified record
Claude Opus 5.5’s most-compared rival is Kimi K3: 5 leads, 1 tie, and 2 not callable across their 8 shared comparisons. Claude Opus 5.5 is priced at $4.00/$20.00 per 1M tokens (in/out) vs Kimi K3’s $3.00/$15.00.
Against the 237 head-to-head comparisons Claude Opus 5.5 shares with other tracked models:
A gap counts for Claude Opus 5.5 only where both sides were run independently and the benchmark still separates models. Losses are listed alongside wins on purpose — Claude Opus 5.5 trails on 0 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 noiseTerminal-Bench 2.1 Terminal ops
±10.6 is noiseAgents' Last Exam Professional work
±3.2 is noiseARC-AGI-2 · high Compositional visual reasoning
±9.2 is noiseNo verdict for Claude Opus 5.5 anywhere on DeepSWE, OSWorld 2.0 · v2 1 full, Toolathlon-Verified ( nothing independently confirmed on both sides).
- None of Claude Opus 5.5’s coding comparisons are independently confirmed on both sides yet.
What changed from Claude Opus 5 to Claude Opus 5.5
The 9 benchmarks both models have been scored on, using the same variant each time. A raw Claude Opus 5.5 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 5 | Claude Opus 5.5 | Change | Verdict |
|---|---|---|---|---|
| HLE | 54.9 | 61.4 | +6.5 | Real gap |
| Terminal-Bench 2.1 | 89.1 | 87.64 | -1.5 | Setup-dependent |
| DeepSWE | 74 | 74.2 | +0.2 | Unverified |
| Toolathlon-Verified | 80.6 | 77.8 | -2.8 | Unverified |
| Agents' Last Exam | 32.2 | 38.2 | +6.0 | Setup-dependent |
| LiveBench | 80.1 | 83.2 | +3.1 | Real gap |
| AnalystAgent | 53.75 | 56.25 | +2.5 | Tie |
| Terminal-Bench 4.0 | 53.54 | 65.15 | +11.6 | Tie |
| OSWorld 2.0 | 77.67 | 81.8 | +4.1 | Unverified |
Claude Opus 5.5 API pricing
$4.00 in / $20.00 out per 1M tokens — official pricing source
What Claude Opus 5.5 costs per job
| Workload | Tokens in / out | Cost |
|---|---|---|
| One long chat turn | 100K / 10K | $0.600 |
| A codebase review | 1,000K / 100K | $6.00 |
| A day of agent work | 10,000K / 1,000K | $60.00 |
Computed from Claude Opus 5.5’s list rates above — cache discounts and batch tiers are not applied.
Claude Opus 5.5 is one of 8 Anthropic models tracked on this site, at these official list prices.
| Model | In / 1M | Out / 1M |
|---|---|---|
| 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 | $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 5.5 benchmark scores
| Benchmark | Score |
|---|---|
| HLE(no tools)[1] Reasoning · ±2 is noise | |
| LiveBench[2] Composite score across 7 domains · ±2.7 is noise | |
| Terminal-Bench 2.1[3] Terminal ops · ±10.6 is noise | |
| Agents' Last Exam[4] Professional work · ±3.2 is noise | |
| ARC-AGI-2(high)[5] Compositional visual reasoning · ±9.2 is noise | |
| AnalystAgent[6] Spreadsheet & document analysis · ±11.2 is noise | |
| Terminal-Bench 4.0(max)[7] Terminal ops · ±12.4 is noise | |
| OSWorld 2.0(v2 1_full)[8] Long-horizon computer use · ±9.7 is noise | |
| DeepSWE[9] Long-horizon coding · ±9.5 is noise | |
| Toolathlon-Verified[10] Multi-tool chores · ±9.7 is noise |
Who ran these numbers: 7 of 10 independent — artificialanalysis.ai (2), vals.ai (2), livebench.ai (1), snorkel.ai (1), arcprize.org (1); vendor self-reported (3).
- HLE: AA's own run at max effort (Adaptive Reasoning, Max Effort, Default Fallback) — text-only 2,158-question subset. Currently the highest HLE score AA has recorded for any model. Anthropic's own max-effort run is 64.4 (system card §8.1.A) — not stored as a row; this cell holds AA's independent run.
- LiveBench: Board row 'Claude 5.5 Opus Thinking (Max Effort)' on the 2026-06-25 LiveBench release; second on the raw board behind Claude Fable 5.1 (Max Effort, 83.4), a 0.2-point gap inside LiveBench's 2.7-point noise band.
- Terminal-Bench 2.1: vals.ai's own Terminus 2 run, board row 'Claude Opus 5.5' (87.64 ± 1.72, pass@1, effort High, adaptive thinking), from its archived Terminal-Bench 2.1 table. Claude Opus 5 and Claude Opus 4.8 served as fallbacks on 26 of 267 tasks (9.74%); counting those as failures lowers it to 79.77.
- Agents' Last Exam: Overall pass rate 38.2, board rank 1 on 2026-10-01 (Claude Code, Max; score 63.2; est. cost $1,340).
- ARC-AGI-2: ARC Prize's own leaderboard row 'Claude Opus 5.5 (High)', dated 2026-09-22: 93.3% at $0.408/task. This model's tiers invert the usual order - High 93.3, XHigh 92.5, Max 91.7 - so the usable ceiling is High, recorded here per this benchmark's convention (the highest populated tier; see the tier caveat in this benchmark's notes, and the Opus 4.8 row whose Max cell is N/A and therefore records High). Corrected 2026-10-03: first recorded as Max 91.7 on the reasoning that Max-to-Max matches the other max-tier rows for cross-model comparability, which contradicted the convention and understated the model by 1.6 points. Read against max-tier rows this is a tier-mismatched comparison by design.
- AnalystAgent: AA's own run, board row 'Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback)': analystAgent 0.5625 = 56.25 pass@5.
- Terminal-Bench 4.0: vals.ai board row — mini-swe-agent harness (single bash tool), pass@1 averaged over 3 full passes, raw value 65.152 at effort max. Artificial Analysis same-conditions run (Max, Default Fallback): 59.6. Official board (grant-funded, Claude Code, max effort, 5-trial pass@1): 64.85 ± 3.11 CI — 0.3 points below this run. Anthropic's launch effort ladder reads Low 38.5 / Med 57.6 / High 64.2 / Xhigh 66.4 / Max 64.8 — the peak sits at xhigh; none of it is stored as a row, this cell holds vals.ai's independent max run.
- OSWorld 2.0: Anthropic launch chart (2026-09-22), labeled OSWorld 2.1, partial reward. System card §8.13.3 gives the conditions: September 10, 2026 task files, assets, and web apps; adaptive thinking at max effort; 1080p; 500 steps; every screenshot retained with Claude API server-side context compaction past 100k tokens; Claude Opus 4.8 as model grader; pass@1 over five runs. Strict pass rate 48.7. The authors' official v2.1 full-set run for the previous generation (Opus 5, max effort) is 77.67 under the official harness — inside the ±9.7 band, so no capability ordering is claimed across the two protocols. Verified on the launch page grid, its embedded data, and the system card. The same card gives a strict pass rate of 48.7 alongside this partial score.
- DeepSWE: System card §8.3: 'an average of 74.2% over five trials' at the standard max-effort configuration.
- Toolathlon-Verified: System card Table 8.14.5.A, 77.8% Pass@1 at adaptive thinking max effort, internal harness against the Toolathlon-Verified June 2026 final release; safety classifiers ON — 1 classifier-stopped and 6 sandbox-escape-monitor-halted trials of 324 counted as fails.
Notes on the record
Pricing is flat — Anthropic's own pricing page states the full 1M-token window bills at the same per-token rate regardless of prompt length, no threshold. $4 input / $20 output per million tokens, cache read at $0.20 (a 0.05x multiplier, lower than Opus 5's 0.1x), cache write $5 (5-minute) or $8 (1-hour), batch API $2/$10, and a research-preview "fast" mode at $8/$40. US-only data residency adds a flat 10% surcharge — not a length tier. Anthropic also claims Opus 5.5 costs 40% less than Opus 5 on typical workloads at default settings, combining the lower price with lower token use — its own test, not independently measured.
Context window 1,000,000 tokens, max output 128,000 (extendable to 300,000 via a Message Batches API beta). Input text and images, output text only. Knowledge cutoff June 2026, one month newer than Opus 5's. License not stated; inferred proprietary from API-only distribution.
Opus 5 became a previous version when this model shipped on 2026-09-22 — same line, same tier, 20% below Opus 5's price — which Anthropic's own model page agrees with, calling Opus 5 a legacy model.
Anthropic's announcement calls Opus 5.5 "the new leading model" that "performs at the level of Claude Fable 5.1 on most work," while Fable 5.1 stays on sale at $10/$50. This site's independent data agrees on the narrow part it can test: Opus 5.5 leads Fable 5.1 on Humanity's Last Exam by 2.3 points, just outside that benchmark's noise band, and the two tie on LiveBench (83.2 vs 83.4). Two more independent rows arrived on 2026-10-01. Agents' Last Exam, where Opus 5.5 tops the board at 38.2 (Max); the tracked Opus 5 row is its 32.2 High run, so the pairing is setup-dependent (at matched Max, 7.3); and vals.ai's TB 2.1 run, 87.64 at high effort. It used Opus 5 and Opus 4.8 as fallbacks on 26 of 267 tasks, which counted as failures reads 79.77.
Compare with
FAQ
Has Claude Opus 5.5 been independently benchmarked?
Yes — seven of the fourteen benchmarks this site tracks carry an independently-run Opus 5.5 score as of 2026-10-08. From launch week: Humanity's Last Exam (61.4%, Artificial Analysis, text-only, max effort, the highest AA has recorded) and LiveBench (83.2, max effort). Recorded 2026-10-01: Snorkel's Agents' Last Exam (38.2%, board rank 1) and vals.ai's Terminal-Bench 2.1 (87.64%, high effort, fallbacks on 26/267 tasks). Added 2026-10-03: ARC-AGI-2 (93.3% at High — this model's tiers invert the usual order, so High is its best) and AA-AnalystAgent (56.25 pass^5). Added 2026-10-08: vals.ai's Terminal-Bench 4.0 (65.15 — the board's top). The vendor-claimed row, OSWorld 2.0's 81.8, is Anthropic's own. GPQA Diamond, SWE-bench Verified, LiveCodeBench, DeepSWE, Toolathlon-Verified and HMMT Feb 2026 have no row here.
Does Claude Opus 5.5 replace Claude Opus 5?
Yes — this site marked Opus 5 a previous version the day 5.5 shipped, 2026-09-22: same line, same tier, newer release, at $4/$20 against Opus 5's $5/$25. Anthropic's own model page calls Opus 5 a legacy model and points here. What it has not done is set a retirement date — the deprecations page still lists Opus 5 "Active", with a tentative retirement no earlier than 2027-07-24 — so Opus 5 stays on sale and keeps its scores. An earlier answer here recorded both statements and picked neither; that left Opus 4.8 and Opus 5 labelled by different rules on identical evidence, and was withdrawn.
Is Claude Opus 5.5 Anthropic's most capable model?
Anthropic doesn't use that phrase, but its announcement calls Opus 5.5 "the new leading model" and says it "performs at the level of Claude Fable 5.1 on most work" — while Fable 5.1 stays on sale at $10/$50 per million tokens, 2.5 times Opus 5.5's $4/$20. On the two benchmarks this site can compare independently, Opus 5.5 leads Fable 5.1 on Humanity's Last Exam by 2.3 points, just past the 2-point noise band, and ties it on LiveBench.
Does Claude Opus 5.5's price change with longer prompts?
No. Anthropic's pricing page states models from Claude 4.6 onward bill the full 1,000,000-token window at one flat rate — a 900,000-token request costs the same per token as a 9,000-token one. The only surcharge is a flat 10% add-on for US-only data residency, which applies regardless of prompt length.
What is Claude Opus 5.5's best verified score?
ARC-AGI-2 at 93.3 (2026-10-03), recorded on a high-effort run — and the tier detail matters: the board's own rows show 92.5 at xHigh and 91.7 at Max for this model, an inversion of the usual order, so 93.3 is the peak row rather than the max-effort row.
What is Claude Opus 5.5's knowledge cutoff?
June 2026, per Anthropic's disclosure, with the full 1,000,000-token window served at a single rate. No long-context price threshold applies to this model.
Further reading
- Models with 10M token context windows 2026 — Claude Opus 5.5 is one of the 47 models it compares.
Benchmark guides
- Humanity's Last Exam — what Claude Opus 5.5’s reasoning score on it does and doesn’t prove.
- LiveBench — what Claude Opus 5.5’s composite score on it does and doesn’t prove.