Claude Sonnet 5: 82.1% SWE-bench & $3/M Pricing (2026)
Claude Sonnet 5 hits 82.1% SWE-bench Verified, first to break 80%, with a 1M-token context window at $3 input / $15 output per 1M tokens, 40% below Opus 4.8.
Claude Sonnet 5 is Anthropic's mid-tier flagship (released June 30, 2026) with a 1M-token context window and 82.1% SWE-bench Verified, the first model to clear 80% on that benchmark. Priced at $3 input / $15 output per 1M tokens (40% below Opus 4.8), it defaults to adaptive thinking and posts 96.2% on GPQA Diamond and 81.2% on OSWorld-Verified computer use.
Claude Sonnet 5, released by Anthropic on June 30, 2026, scores 82.1% on SWE-bench Verified, the first model to break 80%, with a 1M-token context window. It costs $3 input / $15 output per 1M tokens ($2/$10 introductory through Aug 31, 2026), 40% cheaper than Opus 4.8, and leads on GPQA Diamond (96.2%) and computer use (81.2% OSWorld-Verified).
Provider: Anthropic · Family: Claude Sonnet 5
Context window: 1,000,000 tokens · Max output: 128,000
Input modalities: text, image, pdf, tool-calls · Output: text, tool-calls
About Claude Sonnet 5
Claude Sonnet 5 is Anthropic's fifth-generation Sonnet-tier model, released June 30, 2026. Anthropic calls it "the best combination of speed and intelligence" in its lineup, a drop-in upgrade for Sonnet 4.6 that closes the gap to Opus 4.8 while keeping Sonnet's speed and price. It is a dense, hybrid-reasoning Transformer with adaptive thinking on by default; parameter count is undisclosed. It was built to push agentic coding, planning, and browser/terminal tool use down into the mid-price tier that previously required Opus-class models. On benchmarks, Sonnet 5 became the first model to break 80% on SWE-bench Verified at 82.1%, ahead of Gemini 3.1 Pro (80.6%) and GPT-5.4 (~80%). On SWE-bench Pro it scores 63.2% versus Sonnet 4.6's 58.1%, though Opus 4.8 still leads at 69.2%. It posts 96.2% on GPQA Diamond and 84.7% on ARC-AGI-2, and scores 53 on the Artificial Analysis Intelligence Index v4.1, fifth overall behind Claude Fable 5 (60), Opus 4.8 (56), GPT-5.5 xhigh (55), and Opus 4.7 (54). Anthropic has not published MMLU, HumanEval, or GSM8K scores for this release; those legacy benchmarks are treated as saturated in its 2026 reporting in favor of SWE-bench, GPQA, and ARC-AGI. The context window is 1,000,000 tokens, with 128,000 max output tokens on the synchronous Messages API and up to 300,000 on the Batch API via the output-300k-2026-03-24 beta header. Sonnet 5 uses the tokenizer introduced with Opus 4.7: identical text now produces roughly 30% more tokens than under Sonnet 4.6. Reliable knowledge cutoff is January 2026. Modalities are text and image in, text and tool-call out, with no native audio or video I/O. It supports vision, tool calling, structured JSON outputs, prompt caching, web search, and computer use. OSWorld-Verified rose to 81.2% (from 78.5%) and Terminal-Bench 2.1 to 80.4% (from 67.0%), gains Anthropic attributes to better tool selection and error correction in long autonomous loops. Pricing is $3.00 input / $15.00 output per 1M tokens, matching Sonnet 4.6 and 40% below Opus 4.8's $5/$25. Introductory pricing of $2/$10 runs through August 31, 2026 to offset the tokenizer's higher counts. Prompt cache reads run at 0.1x input, and Batch API gets 50% off but is excluded from Zero Data Retention, retained up to 29 days. Deployment spans the Claude API (claude-sonnet-5), AWS, Bedrock, Microsoft Foundry, and Vertex AI (coming soon), plus day-one Claude Code and GitHub Copilot access. Anthropic holds SOC 2 Type I/II, ISO 27001:2022, and ISO/IEC 42001:2023, signs HIPAA BAAs, and offers Zero Data Retention outside the Batch API. Safety partners for the June 30, 2026 system card include METR, Apollo Research, and the UK AI Security Institute. Anthropic reports lower misalignment, hallucination, and sycophancy rates than Sonnet 4.6, plus better prompt-injection resistance. It does not cross the automated AI R&D threshold, shows only limited bio/chem uplift, and is explicitly not optimized for cyber capability, scoring well below Claude Mythos 5 on cyber evaluations. Sonnet 5 suits teams running high-volume agentic coding and computer-use workloads that want near-Opus capability at Sonnet pricing. It fits voice/video products poorly (no native audio/video I/O), breaks configs pinned to non-default sampling parameters (now a 400 error), and is not the choice for dangerous-capability research, where Anthropic steers customers to Opus 4.8 or invite-only Claude Mythos 5.
Pricing
$3 per 1M input tokens, $15 per 1M output tokens at standard pricing, identical to Sonnet 4.6's rate card. Introductory pricing of $2/$10 per 1M tokens applies through August 31, 2026 to offset the ~30% higher token counts from the new tokenizer. Prompt cache reads run at 0.1x the input rate. Batch API gets a 50% discount but is excluded from Zero Data Retention.
Key Features
- Adaptive Thinking: Reasoning mode runs by default on every request unless explicitly disabled, replacing Sonnet 4.6's opt-in extended thinking.
- 1M-Token Context Window: Matches Opus 4.8's context size at a third of the price, with 128K max output on the synchronous API and up to 300K on the Batch API.
- Computer Use: Scores 81.2% on OSWorld-Verified and 80.4% on Terminal-Bench 2.1, both sharp gains over Sonnet 4.6's 78.5% and 67.0%.
- 82.1% SWE-bench Verified: First model to break the 80% ceiling on SWE-bench Verified, ahead of Gemini 3.1 Pro and GPT-5.4.
- Prompt Caching: Cache reads run at 0.1x the standard input rate, cutting costs sharply on repeat-context agentic workloads.
Pros
- First model to clear 80% SWE-bench Verified at 82.1%, beating Gemini 3.1 Pro and GPT-5.4.
- 40% cheaper than Opus 4.8 on both input and output ($3/$15 vs $5/$25 per 1M tokens), with $2/$10 introductory pricing through August 31, 2026.
- Computer use and terminal automation jumped sharply: 81.2% OSWorld-Verified and 80.4% Terminal-Bench 2.1, up from 78.5% and 67.0% on Sonnet 4.6.
Cons
- No native audio or video input/output, so voice and video apps still need a separate model in the stack.
- New tokenizer inflates token counts roughly 30% versus Sonnet 4.6, requiring max_tokens and cost budgets to be re-tuned on migration.
- Sampling params (temperature, top_p, top_k) now error at non-default values, breaking configs carried over directly from Sonnet 4.6.
Benchmarks
- arc agi 2: 84.7
- gpqa diamond: 96.2
- swe bench verified: 82.1
- artificial analysis intelligence index: 53
Frequently Asked Questions
What is Claude Sonnet 5 and who built it?
Claude Sonnet 5 is Anthropic's fifth-generation Sonnet-tier large language model, released June 30, 2026, roughly nine months after Sonnet 4.5 and three months after Sonnet 4.6. It is a dense, hybrid-reasoning Transformer with adaptive thinking on by default; Anthropic has not disclosed a parameter count. Anthropic describes it as "the best combination of speed and intelligence" in its lineup, designed to bring near-Opus agentic coding and computer-use performance down to Sonnet's price point. It scores 82.1% on SWE-bench Verified, the first model to break 80% on that benchmark, and 96.2% on GPQA Diamond. It sits below Claude Opus 4.8 and above Claude Haiku 4.5 in Anthropic's current lineup, and was built specifically to beat GPT-5.4 and Gemini 3.1 Pro on agentic coding and computer-use tasks. It launched with a 1M-token context window at $3 input / $15 output per 1M tokens.
How much does Claude Sonnet 5 cost per 1M tokens?
Claude Sonnet 5 costs $3.00 per 1M input tokens and $15.00 per 1M output tokens at standard pricing, identical to Sonnet 4.6's rate card. Anthropic is running introductory pricing of $2.00 input / $10.00 output per 1M tokens through August 31, 2026 to offset the new tokenizer's roughly 30% higher token counts, after which standard pricing applies. Prompt cache reads run at 0.1x the input rate ($0.30 per 1M at standard pricing). The Batch API gets a 50% discount ($1.50/$7.50 per 1M) but is excluded from Zero Data Retention and retains data up to 29 days. A daily agentic coding loop processing 1M input and 200K output tokens costs roughly $6.00 at standard pricing. This is 40% cheaper than Claude Opus 4.8's $5/$25 per 1M on both input and output. There is no self-hosting option since the model is proprietary and API-only.
What is Claude Sonnet 5's context window and max output?
Claude Sonnet 5 has a 1,000,000 token context window, matching Claude Opus 4.8 and shared across Anthropic's current top-tier models. Max output on the synchronous Messages API is 128,000 tokens, and the Message Batches API supports up to 300,000 output tokens using the output-300k-2026-03-24 beta header. The model uses the same new tokenizer introduced with Opus 4.7, which produces roughly 30% more tokens for identical input text than Sonnet 4.6's tokenizer, so raw token counts from the prior generation should be recounted rather than reused. This is a substantial jump from Sonnet 4.5's 200K context window. Long-context recall is reported as strong given the model's high scores on agentic, multi-step benchmarks that depend on maintaining state across long tool-use sessions. Document handling covers PDF and image input alongside text within that same 1M budget.
How does Claude Sonnet 5 compare on benchmarks vs GPT-5.4 and Gemini 3.1 Pro?
Claude Sonnet 5 scores 82.1% on SWE-bench Verified, ahead of Gemini 3.1 Pro (80.6%) and GPT-5.4 (approximately 80%), making it the first model to clear the 80% mark on that benchmark. On the harder SWE-bench Pro variant it scores 63.2%, up from Sonnet 4.6's 58.1%, though Claude Opus 4.8 still leads at 69.2% on that specific test. Sonnet 5 posts 96.2% on GPQA Diamond and 84.7% on ARC-AGI-2, both reported as leading its comparison set. On the Artificial Analysis Intelligence Index v4.1, it scores 53 at max effort, placing fifth overall behind Claude Fable 5 (60), Opus 4.8 (56), GPT-5.5 xhigh (55), and Opus 4.7 (54). The practical takeaway of the SWE-bench gap is that Sonnet 5 now closes most of the coding-quality gap to Opus-tier models while costing 40% less per token. Anthropic has not published an MMLU-Pro or AIME 2025 score for Sonnet 5 specifically.
Is Claude Sonnet 5 open source or proprietary?
Claude Sonnet 5 is fully proprietary, API-only, with no downloadable weights and no open license of any kind. It is accessible through the direct Anthropic Claude API (model ID claude-sonnet-5), Claude Platform on AWS, Amazon Bedrock (anthropic.claude-sonnet-5, both global and regional endpoints), Microsoft Foundry on Azure, and Google Vertex AI, which Anthropic listed as coming soon at launch. It also ships inside Claude Code and became generally available for GitHub Copilot on release day. Anthropic's commercial terms govern all usage; there is no research-only or non-commercial carve-out. Because the weights are closed, self-hosting, air-gapped deployment, and fine-tuning the base model are not available, unlike open-weights models such as Llama 4 or Qwen3.
What modalities does Claude Sonnet 5 support?
Claude Sonnet 5 accepts text, image, and PDF input and produces text and tool-call output; it has no native audio or video input or output in either direction. It supports native function calling and tool use, structured JSON outputs, prompt caching, web search, and computer use (screen-reading and control for desktop automation). Computer-use performance jumped sharply versus Sonnet 4.6: 81.2% on OSWorld-Verified (up from 78.5%) and 80.4% on Terminal-Bench 2.1 (up from 67.0%). Adaptive thinking is on by default for every request rather than an opt-in extended-thinking mode, a change from Sonnet 4.6's behavior. Teams building voice or video applications will need to pair Sonnet 5 with a separate ASR/TTS or video model, since neither modality is natively supported.
Does Claude Sonnet 5 train on user data?
No. Anthropic does not train on API inputs or outputs by default. Standard API data is retained according to Anthropic's normal retention policy, and Zero Data Retention is available for eligible customers, meaning inputs are not stored at rest after the response is returned except where required by law or to combat misuse. The Message Batches API is the one exception: it is not eligible for Zero Data Retention and retains request and response data for up to 29 days. Anthropic holds SOC 2 Type I and Type II, ISO 27001:2022, and ISO/IEC 42001:2023 certifications, and signs HIPAA Business Associate Agreements covering the first-party API and Enterprise plans for organizations handling protected health information. The EU AI Act classifies Sonnet 5 as a general-purpose AI model with systemic risk obligations given its capability level and reach.
Who is Claude Sonnet 5 best for and who should avoid it?
Claude Sonnet 5 is best for teams running high-volume agentic coding loops that need near-Opus SWE-bench performance without Opus pricing, teams automating computer-use and terminal workflows given its jump to 81.2% OSWorld-Verified and 80.4% Terminal-Bench 2.1, and enterprise teams that need SOC 2, HIPAA-ready, and Zero Data Retention options at Sonnet-tier cost. It also suits anyone migrating off Sonnet 4.6 who wants a same-price upgrade with a real capability jump. It is a poor fit for voice or video-first products since it has no native audio or video I/O, for teams reusing exact Sonnet 4.6 sampling configs since non-default temperature, top_p, or top_k now return a 400 error, and for dangerous-capability-adjacent cyber research, where Anthropic explicitly states Sonnet 5 underperforms Claude Mythos 5 by design. Teams needing the absolute highest coding ceiling on SWE-bench Pro should still consider Opus 4.8, which leads that specific benchmark at 69.2% versus Sonnet 5's 63.2%.