AnthropicPrevious version · Claude Sonnet 5.5

Claude Sonnet 5

Sonnet 5's $2/$10 launch pricing was introductory, scheduled to rise to $3/$15 on 2026-09-01 — Anthropic cancelled that increase, so $2/$10 is now permanent. Ten of twelve benchmark results here are independent third-party runs; the exceptions are Anthropic's own OSWorld and HLE-with-tools figures — but Sonnet 5's newer tokenizer produces roughly 30% more tokens for the same text, so the effective cost-per-task falls by less than the sticker price suggests.

Claude Sonnet 5 benchmarks and pricing, every number sourced: 12 tracked Claude Sonnet 5 benchmark scores (10 independently run, 2 still resting on a vendor’s own claim), priced at $2.00 per million input tokens and $10.00 per million output.

Released
2026-06-30
License
proprietary
Context window
1M tokens
Knowledge cutoff
2026-01
Verified
sources checked 2026-08-17–2026-10-25
Parameters
Not disclosed
Architecture
Not disclosed

Claude Sonnet 5’s verified record

Claude Sonnet 5’s featured comparison is Claude Opus 5: 3 trails, 1 tie, and 7 not callable. Claude Sonnet 5 is priced at $2.00/$10.00 per 1M tokens (in/out) vs Claude Opus 5’s $5.00/$25.00. Full verdict →

Against the 297 head-to-head comparisons Claude Sonnet 5 shares with other tracked models:

66 real gaps48 inside the band183 not called

A gap counts for Claude Sonnet 5 only where both sides were run independently and the benchmark still separates models. Losses are listed alongside wins on purpose — Claude Sonnet 5 trails on 48 of them.

HLE · no tools Reasoning

±2 is noise
Behind
DeepSeek V4.1 Flash −2.0 · Claude Haiku 5.5 −3.1 · Step 5 Preview −5.2 · Kimi K3 −5.6 · Gemini 3.1 Pro Preview −5.7 · MiniMax M3.1 Flash −6.7 · Muse Spark 1.3 −7.4 · MiMo-V2.6-Pro −8.1 · Muse Wonder −8.8 · Kimi K3.1 Pro −11.2 · GPT-6.1 Sol −11.6 · GPT-6 Astra −13.4 · Claude Sonnet 5.5 −13.7 · Claude Opus 5.5 −20.1 · 6 previous: GPT-6 Sol −6.6 · Claude Opus 4.8 −7.4 · GPT-5.6 Sol −8.2 · Claude Opus 5 −13.6 · Claude Fable 5 −14.2 · Claude Fable 5.1 −17.8
Ahead
MiniMax M3 +2.3 · GPT-6 Luna +2.8 · Qwen3.8-Flash-Next +3.3 · Mistral Large 4 +6.3 · Beam 1.0 +6.6 · Nemotron 3.5 Super VL +10.6 · 1 previous: DeepSeek V4 Flash (0731) +2.7
Tie
6 models within ±2
Unverified
1 model — vendor-reported on one side
Setup-dependent
7 models — scored on a different harness or effort tier

Terminal-Bench 4.0 · max Terminal ops

±12.4 is noise
Behind
Muse Spark 1.3 −15.2 · Claude Haiku 5.5 −25.8 · GLM-5.3 −29.3 · GPT-6.1 Sol −45.4 · GPT-6 Astra −50.0 · Claude Sonnet 5.5 −54.5 · Claude Opus 5.5 −55.6 · 6 previous: Claude Opus 4.8 −13.6 · GPT-5.6 Sol −28.3 · Claude Fable 5 −31.8 · GPT-6 Sol −34.8 · Claude Opus 5 −43.9 · Claude Fable 5.1 −48.5
Tie
4 models within ±12.4
Unverified
1 model — vendor-reported on one side
Setup-dependent
1 model — scored on a different harness or effort tier

DeepSWE Long-horizon coding

±9.5 is noise
Behind
Kimi K3 −15.0 · GLM-5.3 −15.0 · Gemini 3.8 Flash −20.0 · GPT-6 Astra −20.0 · 6 previous: Gemini 3.7 Flash −11.0 · Grok 4.6 −13.0 · GPT-5.6 Luna −13.0 · Claude Fable 5 −16.0 · GPT-5.6 Sol −19.0 · Claude Opus 5 −20.0
Ahead
1 previous: GLM-5.2 +10.0
Tie
5 models within ±9.5
Unverified
19 models — vendor-reported on one side
Setup-dependent
2 models — scored on a different harness or effort tier

AnalystAgent Spreadsheet & document analysis

±11.2 is noise
Behind
Claude Sonnet 5.5 −11.3 · 1 previous: Claude Fable 5.1 −11.3
Ahead
Tie
16 models within ±11.2
Setup-dependent
5 models — scored on a different harness or effort tier

LiveBench Composite score across 7 domains

±2.7 is noise
Behind
Kimi K3 −3.2 · Muse Spark 1.3 −5.6
Ahead
Claude Haiku 5.5 +3.9 · GLM-5.3-Flash +4.4 · MiniMax M3 +8.7 · 1 previous: GLM-5.2 +2.8
Tie
9 models within ±2.7
Setup-dependent
14 models — scored on a different harness or effort tier

Toolathlon-Verified Multi-tool chores

±9.7 is noise
Ahead
Gemini 3.1 Pro Preview +10.5 · 1 previous: GLM-5.2 +11.7
Tie
6 models within ±9.7
Unverified
11 models — vendor-reported on one side

Terminal-Bench 2.1 Terminal ops

±10.6 is noise
Behind
GPT-6 Astra −12.8
Tie
2 models within ±10.6
Unverified
3 models — vendor-reported on one side
Setup-dependent
29 models — scored on a different harness or effort tier

No verdict for Claude Sonnet 5 anywhere on GPQA Diamond, LiveCodeBench, SWE-bench Verified ( saturated); HLE · with tools, OSWorld 2.0 · v2 1 full ( nothing independently confirmed on both sides).

Claude Sonnet 5 API pricing

$2.00 in / $10.00 out per 1M tokens — official pricing source

What Claude Sonnet 5 costs per job

Claude Sonnet 5 cost for three reference workloads, computed from its list rates
WorkloadTokens in / outCost
One long chat turn100K / 10K$0.300
A codebase review1,000K / 100K$3.00
A day of agent work10,000K / 1,000K$30.00

Computed from Claude Sonnet 5’s list rates above — cache discounts and batch tiers are not applied.

Claude Sonnet 5 is one of 9 Anthropic models tracked on this site, at these official list prices.

Anthropic model pricing, official list rates
ModelIn / 1MOut / 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 Sonnet 5 benchmark scores

Claude Sonnet 5 benchmark scores, provenance, and source links
BenchmarkScore
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
GPQA Diamondsaturated[5]
Expert science Q&A — not ranked at any gap size
LiveCodeBenchsaturated[6]
Contest coding — not ranked at any gap size
Toolathlon-Verified[7]
Multi-tool chores · ±9.7 is noise
LiveBench[8]
Composite score across 7 domains · ±2.7 is noise
AnalystAgent[9]
Spreadsheet & document analysis · ±11.2 is noise
Terminal-Bench 4.0(max)[10]
Terminal ops · ±12.4 is noise
OSWorld 2.0(v2 1_full)[11]
Long-horizon computer use · ±9.7 is noise
HLE(with tools)[12]
Reasoning · ±2 is noise

Who ran these numbers: 10 of 12 independent — vals.ai (4), artificialanalysis.ai (2), tbench.ai (1), deepswe.datacurve.ai (1), toolathlon.xyz (1), livebench.ai (1); vendor self-reported (2).

  1. HLE: Artificial Analysis's own run (max effort, text-only, no tools; 2,158 questions, pass@1). Not comparable to vendors' with-tools HLE claims. Corrected 2026-10-03: previously cited the evaluations URL with ?models=claude-sonnet-5, a client-side-only parameter whose response is byte-identical to the unfiltered page; the value is now cited from AA's model page (raw 0.412882).
  2. Terminal-Bench 2.1: Sonnet 5's Terminal-Bench 2.1 row was re-verified 2026-09-03 in a browser after npm run check-sources flagged it: tbench.ai rebuilt the site around 2026-08-27 (new default board Terminal-Bench 4.0, version now a client-side selection on a single URL), so the old deep link no longer serves 2.1 in static HTML — see the claude-opus-4-8 row on this same benchmark for the full mechanism. Value confirmed unchanged: 74.6%, rank 10 of 17, 'Sonnet 5 (high)' via Claude Code (CI now shown as ±3.2%, not ±1.6 — recomputed, not a re-run). Cross-corroborated by vals.ai's independent run at 74.53%. Anthropic's card reports 80.4 at xhigh via the mini-SWE-agent harness (chosen over Terminus-2 for timeout robustness) — a different harness from the tbench.ai official run stored here, and not stored as a row.
  3. DeepSWE: 54%±4 pass@1 on the mini-swe-agent harness, rank 13/17; $26.40 avg cost/task. Board updated 2026-08-13.
  4. SWE-bench Verified: 79.60%±1.80, rank 22/83 (updated 2026-08-14). vals.ai's earlier Jun-30 setup gave 75.49 — harness changed since, numbers not directly comparable across dates. Anthropic's card reports 85.2 on SWE-bench Verified, a 500-problem subset — not stored as a row; this cell holds vals.ai's run.
  5. GPQA Diamond: 88.89%±2.22, rank 30/133 (updated 2026-08-15). vals.ai notes the benchmark is largely saturated — 24 models at 90%+.
  6. LiveCodeBench: 82.43%±1.09 on v6, rank 50/138 (updated 2026-08-15).
  7. Toolathlon-Verified: Pass@1 71.6±1.2 on the current Toolathlon-Verified table (Pass@3 83.3, Pass^3 53.7 also shown but not tracked here). Anthropic's own numbers: 54.3 Pass@1 on the card's older internal harness (pinned offline snapshot, ~3 points above a strictly upstream-equivalent setup, per the card) and 74.7 on the Verified-harness re-run in the later Opus 5.5 / Sonnet 5.5 cards — neither stored as a row; this cell holds the official board's 71.6.
  8. LiveBench: Board row "Claude Sonnet 5 xHigh Effort" on the 2026-06-25 LiveBench release.
  9. AnalystAgent: AA's own run, board row 'Claude Sonnet 5 (Adaptive Reasoning, Max Effort)'; the board rounds to one decimal. pass@1 61.5, pass@5 77.5. Source re-pointed 2026-09-30: AA's AA-AnalystAgent leaderboard page no longer embeds this model in its default selection; its model page still shows 46.25%.
  10. Terminal-Bench 4.0: vals.ai board row — mini-swe-agent harness (single bash tool), pass@1 averaged over 3 full passes, raw value 9.596 at effort max. Official board (grant-funded, Claude Code, max effort, 5-trial pass@1): 12.42 ± 3.06 CI — 2.8 points above this run. Maintainer-disclosed execution caveat: Sonnet 5 hit timeouts and output-token-exceeded errors on its official run (21.6B tokens versus Opus 5's 6.5B).
  11. OSWorld 2.0: Anthropic's rerun of Sonnet 5 under the same v2.1 configuration, published on the Sonnet 5.5 launch chart (2026-09-28). Partial reward 57.0, strict pass rate 25.6. No official-board row exists for this model on any release, so this is the only OSWorld figure in the record for it — and it is vendor-run under the modified protocol. Verified on the launch page grid, its embedded data, and the Sonnet 5 system card text.
  12. HLE: System card Table 8.1.A at the standard max-effort configuration. The later Sonnet 5.5 card re-lists Sonnet 5 at 54.9 with no explanation — the contemporaneous card's number is stored.

Notes on the record

Mid-tier of the Claude 5 family; Anthropic's docs place its predecessor Sonnet 4.6 in 'Legacy models' (platform.claude.com/docs/en/about-claude/models/overview). A previous version since Claude Sonnet 5.5 shipped on 2026-09-28: this site's rule is same product line and price tier, newer release, and the price is unchanged at $2/$10. Anthropic's own model page agrees from its side, calling Sonnet 5 "a legacy model" and naming 5.5 the current Sonnet (checked 2026-10-04). What Anthropic has not done is set a retirement date — its deprecations page still lists Sonnet 5 "Active", a status that page's own glossary defines as "fully supported and recommended for use", with a tentative retirement no earlier than 2027-06-30. Sonnet 5 also serves as the visible fallback for Sonnet 5.5's cyber and frontier-LLM safety blocks (checked 2026-09-29). Launch's introductory $2/$10 pricing, originally scheduled to rise to $3/$15 on 2026-09-01, was instead made the permanent standard price on 2026-08-10 — the increase will not occur, per the official pricing docs (platform.claude.com/docs/en/about-claude/pricing) and Anthropic's own announcement (x.com/claudeai/status/2086891169217122586).

Adaptive thinking with an effort control (low/medium/high/xhigh/max) that defaults to high; unlike Sonnet 4.6, the old manual extended-thinking mode (thinking: {type: "enabled", budget_tokens: N}) and non-default sampling parameters now return a 400 error — thinking can still be disabled outright via thinking: {type: "disabled"} — and Priority Tier (Sonnet 4.6) is not offered on Sonnet 5.

Max output 128k (300k via the Batch API extended-output beta). The full 1M-token context window bills at standard per-token rates with no long-context surcharge or threshold tier; the newer tokenizer used from Sonnet 5 onward produces roughly 30% more tokens than Sonnet 4.6's for the same text (confirmed on Anthropic's what's-new-sonnet-5 docs), so per-task cost savings are smaller in practice than the per-token price cut alone suggests.

The card's OSWorld-Verified 81.2 predates this site's versioned slices; the stored 57 is the Sonnet 5.5 card's later v2.1 re-eval.

Compare with

FAQ

Has Claude Sonnet 5 been independently benchmarked?

Yes — every score tracked for Claude Sonnet 5 on this page (HLE, Terminal-Bench 2.1, DeepSWE, SWE-bench Verified, GPQA Diamond, LiveCodeBench, Toolathlon-Verified, LiveBench, and AA-AnalystAgent) is an independent third-party run — from ArtificialAnalysis, tbench.ai, the DeepSWE board, Vals AI, Toolathlon's own board, or LiveBench's own board — observed 2026-08-17 through 2026-09-29, rather than a number self-reported by Anthropic. That's 10 of 12 listed benchmarks independently run; the self-reported entries are Anthropic's OSWorld 2.0 number, added 2026-10-08, and its HLE with-tools number from the original system card.

Why is Claude Sonnet 5 marked as a previous version?

Because Claude Sonnet 5.5 shipped on 2026-09-28, in the same line and at the same price tier — which is how this site decides, not what a vendor says. Anthropic's own page backs it from its side too, calling Sonnet 5 a legacy model and naming 5.5 the current Sonnet. What Anthropic has not done is set a retirement date: its deprecations page still lists Sonnet 5 "Active", which that page's glossary defines as fully supported and recommended for use, with a tentative retirement no earlier than 2027-06-30. Sonnet 5 therefore stays on sale at $2/$10 and keeps every score on this page — previous version means newer in the lineup, not switched off.

Will Claude Sonnet 5's $2/$10 pricing go up later?

No. The $2-per-million-input / $10-per-million-output rate launched as introductory pricing with a scheduled rise to $3/$15 set for 2026-09-01, but Anthropic canceled that increase on 2026-08-10 and made the launch rate the permanent standard price, per the official pricing docs.

What does the effort setting do on Claude Sonnet 5, and can its thinking be turned off?

Effort (low, medium, high, xhigh, or max) controls how many tokens Claude Sonnet 5 spends reasoning before it answers, and it defaults to high. Thinking itself can still be turned off by passing thinking: {type: "disabled"} — what Anthropic actually removed is the old manual extended-thinking mode (thinking: {type: "enabled", budget_tokens: N}), which now returns a 400 error, per Anthropic's docs.

Is Claude Sonnet 5 a drop-in replacement for Claude Sonnet 4.6?

Mostly. Per Anthropic's docs, Claude Sonnet 5 has three behavior changes from Sonnet 4.6: adaptive thinking now runs by default (it can still be turned off via thinking: {type: "disabled"}), the old manual extended-thinking mode is removed and returns a 400 error, and non-default sampling parameters (temperature, top_p, top_k) also now return a 400 error instead of being silently accepted. Separately, Priority Tier — available on Sonnet 4.6 — is not offered on Sonnet 5.

Is Claude Sonnet 5 cheaper than Claude Opus 5?

Per Anthropic's pricing docs, yes: Sonnet 5 lists at $2/$10 per million input/output tokens versus Opus 5's $5/$25, less than half the per-token price. Because Sonnet 5's tokenizer produces roughly 30% more tokens than Sonnet 4.6's for the same text, the real cost gap on a given task will run narrower than the sticker prices alone suggest.

How much can prompt caching or batch processing cut Claude Sonnet 5's API costs?

Cache reads cost 0.1x the base input price ($0.20 per million tokens), and the Batch API cuts both input and output by 50% ($1/$5 per million tokens), so a cached, batched workload can run well below the $2/$10 sticker price. Routing to US-only inference instead adds a 1.1x multiplier, per Anthropic's pricing docs.

Further reading

Benchmark guides