benchrHead-to-head: for the case-by-case breakdown of what the extra dollar buys over Sonnet 4.6, and where Opus 4.8 still wins, see the sections below. The launch coverage has the fuller capability picture.
Pricing breakdown
| Tier | Rate / 1M tokens |
|---|---|
| Intro input (through Aug. 31) | $2.00 |
| Intro output (through Aug. 31) | $10.00 |
| Standard input (from Sep. 1) | $3.00 |
| Standard output (from Sep. 1) | $15.00 |
| Cached input | $0.20 intro / $0.30 standard |
| Batch (50% off) | $1.00 / $5.00 intro; $1.50 / $7.50 standard |
| Context window | 1,000,000 tokens |
| Max output | 128,000 tokens |
A new price tier, not a new number after "Sonnet"
Claude Sonnet 5 is Anthropic's second model built on the Mythos-class architecture it introduced with Claude Fable 5 on June 9. Rather than call this "Sonnet 4.7" or "4.8", Anthropic rolled the new architecture down to the Sonnet price tier. During launch, it is even cheaper: $2/1M input and $10/1M output through August 31, 2026. Starting September 1 it moves to $3/$15, matching Sonnet 4.6 and staying below Opus 4.8's $5/$25. Cached input runs $0.20/1M during intro and $0.30 afterward, and the batch API cuts both rates in half.
Outscoring Opus 4.8 on one published benchmark
Claude Sonnet 5 is reported at 89.4% on SWE-bench Verified, above Claude Opus 4.8's 88.6% in the cited snapshot. SWE-bench is a useful, defined software-engineering benchmark; it is not proven to be the single best predictor of production coding-agent success. The published card also lists 71.8% on SWE-bench Pro, 85.6% on Terminal-Bench 2.1, 96.0% on HumanEval, and 93.5% on MATH. Opus 4.8 is listed higher on GPQA Diamond, 93.6% versus 92.0%. Test the task mix you run rather than treating either score as a universal routing rule.
The rest of the scorecard: LMSYS Arena 1435, MMLU 93.8%, ARC-AGI-2 20.0, and 42.5% on Humanity's Last Exam without tools — a strong general profile for a model priced below the flagship.
The 128K maximum output limit
Claude Sonnet 5 supports 128,000 output tokens per response — twice Sonnet 4.6's 64,000-token ceiling. For workloads that generate long documents, large diffs, or multi-file code changes in a single call, this removes a real constraint: tasks that used to require splitting a response across multiple Sonnet 4.6 calls now fit in one Sonnet 5 call, with the full 1,000,000-token context available to read from.
Adaptive thinking and explicit safety refusals
Like Fable 5, Claude Sonnet 5 inherits adaptive thinking: the model decides its reasoning depth per request, and there is no manual extended-thinking toggle. Safety-classified offensive-cyber, biology/chemistry, and distillation requests return an explicit refusal. No Opus 4.8 request is made by default; another-model retry and its billing occur only when an application or product configures that behavior.
Cost scenarios
A 928K-token session (800K input, 128K output — a single call that uses the full max-output ceiling): 800,000/1M × $2 + 128,000/1M × $10 = $1.60 + $1.28 = $2.88 during the introductory window. Route the same session through the batch API and it drops to $0.80 + $0.64 = $1.44, exactly half. From September 1, the same standard-rate session is $4.32.
Add caching: if a stable system prompt or repo snapshot covers 720K of that 800K input (90% cache hit), the input leg becomes 720,000/1M × $0.20 + 80,000/1M × $2 = $0.144 + $0.16 = $0.304, for a session total of $0.304 + $1.28 = $1.58 during intro — a 45% reduction from the uncached $2.88.
A typical day of usage: a coding-agent team running 25 of those 928K-token sessions a day pays 25 × $2.88 = $72/day uncached during intro, or 25 × $1.44 = $36/day on batch. At a smaller, monthly scale — 20M input plus 5M output tokens — Sonnet 5 costs 20 × $2 + 5 × $10 = $40 + $50 = $90/month through August 31. From September 1, the same workload is $135/month, matching Sonnet 4.6; Opus 4.8 is $225/month.
Use-case fit
Candidate for: Sonnet 4.6 users evaluating a capability step without Opus pricing, coding-agent pipelines that can use the 128K max output ceiling, and teams that want to test whether the reported SWE-bench difference appears on their own tasks.
Compare alternatives if: your workload emphasizes scientific reasoning, your traffic is high-volume and routine, or another current model meets your pass threshold at lower total cost. Benchmark differences do not replace a matched workload evaluation.
Decision checklist
Coming from Sonnet 4.6: run your hardest coding-task eval on both. If Sonnet 5's SWE-bench-level gains show up in your results, the 33% price increase is easy to justify against the 128K output ceiling alone.
Comparing against Opus 4.8: if your tasks are coding-dominant, Sonnet 5 is both cheaper and higher-scoring on SWE-bench Verified. If they're reasoning-heavy in the GPQA sense, Opus 4.8's 93.6% still buys something Sonnet 5 doesn't match.