Claude Sonnet 4.6: Full Spec Sheet (June 2026)

Sonnet 4.6 is Anthropic's production-tier model: not as expensive as Opus, not as small as Haiku, with the full Claude 4 feature set. Anthropic positions Sonnet as the right choice for 'most production workloads' and the data backs that up — across customer deployments, Sonnet handles the bulk of traffic with Opus reserved for the hard subset and Haiku reserved for the cheap subset.

Same architecture as Opus 4.7 (same Claude 4 base), trained on the same data with the same RLHF discipline, with a smaller model size that trades a few percentage points of quality on hard reasoning for a 5× price reduction. The feature surface is identical: tool use, parallel tool calls, extended thinking, prompt caching, vision input, structured outputs via tool schemas, the Batch API.

Sonnet's standout feature vs every other mid-tier model: the optional 1M-token context window. Opt-in via the `anthropic-beta: context-1m-2025-08-07` header. Anthropic doubles input price and adds 50% to output for the portion of a request that exceeds 200K tokens — but the long-context option lets Sonnet compete with Gemini 2.5 Pro on document-scale workloads. No other Claude tier exposes 1M.

Standard rates: `cost = (input_tokens / 1M) × $3 + (output_tokens / 1M) × $15`. The representative 1,000-in / 500-out call: `0.001 × $3 + 0.0005 × $15 = $0.003 + $0.0075 = $0.0105`. About 1¢ per call — 5× cheaper than Opus on the same tokens, but 8× more expensive than gpt-5-mini.

Prompt caching is the largest cost lever. Mark stable blocks with `cache_control: {type: 'ephemeral'}` and Anthropic caches that prefix for 5 minutes (or 1 hour with `ttl: '1h'`). Cached reads drop from $3/M to $0.30/M — 90% off. First-write premiums: 25% on 5-min TTL, 100% on 1-hour TTL.

Worked: a customer-support pipeline with a 3,000-token cached system prompt, 100K calls/month, 600 dynamic input + 200 output. Without caching: `(0.0036 × $3 + 0.0002 × $15) × 100K = $4,080/month`. With caching (90% hit rate): system prefix bills at $0.30/M most of the time, total drops to ~$1,250/month. ~70% off — entirely from prompt structure, no model change.

Batch API on top: 50% off both streams for asynchronous workloads. The stack of caching + batching turns Sonnet into a price-competitive option even against gpt-5-mini for the right workload shape. Worked $ across the Claude family: Claude API cost calculator.

Opt-in to 1M context by adding `anthropic-beta: context-1m-2025-08-07` to your API request headers. Sonnet then accepts up to 1,000,000 input tokens in a single call. The pricing changes when input exceeds 200K: input price tier shifts to $6/M (2×), output to $22.50/M (1.5×).

Worth the premium when: you need to fit an entire long document (a full book, a complete codebase chunk, a months-long conversation history) in one call for cross-section reasoning. The 1M context lets Sonnet do things no 200K model can — answer questions that require pulling from page 2 and page 600 of the same document.

Not worth the premium when: retrieval-augmented generation (RAG) over chunked documents is sufficient. RAG on Sonnet at standard 200K is dramatically cheaper than 1M-context Sonnet, and on most knowledge-Q&A workloads, well-tuned RAG hits the same answer quality.

Comparison: Gemini 2.5 Pro also offers 1M context at a different pricing structure ($2.50 input >200K vs Sonnet's $6). For raw 1M-context cost, Gemini 2.5 Pro is cheaper. For Anthropic-quality voice + discipline at 1M, Sonnet 4.6 1M beta is the only option.

Sonnet 4.6 supports the same extended thinking feature as Opus 4.7. Configure via `thinking: {type: 'enabled', budget_tokens: 3000}` in the API call. Sonnet will burn up to 3,000 internal reasoning tokens before producing the visible answer.

The cost dynamic is different on Sonnet than on Opus because Sonnet's output rate is 5× cheaper. A 3,000-token thinking budget on Sonnet adds `0.003 × $15 = $0.045` to the call. The same budget on Opus adds `0.003 × $75 = $0.225`. Extended thinking is more economically defensible on Sonnet — you can afford to apply it to a wider class of tasks.

Typical Sonnet thinking budgets: 1,000-2,000 tokens for routine analysis tasks, 3,000-5,000 for code synthesis with non-trivial logic, 5,000+ for math/proof tasks. As on Opus, don't add thinking to classification or extraction — it adds cost without improving quality.

**Pick Sonnet 4.6** for the production default in any workflow that benefits from Anthropic's voice, discipline, or tooling: customer-facing chat, content generation, structured-data pipelines, summarization with quality requirements, code review and explanation, agentic workflows that don't need Opus-tier planning. Most teams that standardize on Claude run >80% of traffic on Sonnet.

**Pick Opus 4.7** when Sonnet's quality on a specific hard subset isn't enough — complex agentic loops, multi-file code synthesis with strict correctness, legal/financial analysis where one wrong answer is costly, deep research synthesis. Pay 5× for the Opus boost only on the tasks that need it.

**Pick gpt-5-mini** ($0.25 / $2 per 1M) over Sonnet when: cost is the dominant constraint, the task is mechanical (classification, extraction), you're already in the OpenAI ecosystem, or you need the larger 400K context without the 1M-beta complexity. gpt-5-mini is 12× cheaper than Sonnet on input. Sonnet beats it on voice, long-form writing, and instruction-following discipline.

Cross-tier head-to-head: Claude Sonnet vs GPT-5 mini.

Sonnet 4.6 ships with Anthropic's full tool-use API: define tools as JSON Schema in the `tools` parameter, Sonnet picks one (or several, in parallel) and returns the arguments in a `tool_use` content block. Parallel tool use is on by default; opt out with `disable_parallel_tool_use: true`.

Structured outputs follow the same tool-use pattern as Opus: define a tool whose input schema is your desired output schema, force the call with `tool_choice: {type: 'tool', name: 'extract_data'}`. The JSON Schema enforcement is reliable; invalid outputs are extremely rare on well-formed schemas.

XML-tag prompts work well: `<task>...</task>`, `<context>...</context>`, `<example>...</example>`, `<output_format>...</output_format>`. Sonnet, like Opus, is trained to attend to these reliably and Anthropic's advanced-pattern docs use them consistently. Markdown-headed or plain-paragraph prompts also work but tend to underperform vs XML-tagged equivalents on complex multi-section instructions.

Every number on this page was verified against Anthropic's live documentation on 2026-06-20. Sources: docs.anthropic.com/en/docs/about-claude/models for context window, modalities, and feature support; anthropic.com/pricing for input/output/cached prices and the 1M-beta tier; docs.anthropic.com/en/docs/build-with-claude/prompt-caching for cache write/read mechanics.

The 1M context beta requires the explicit `anthropic-beta: context-1m-2025-08-07` header. Anthropic versions beta features by ID — when the beta moves to general availability, the header changes. Watch docs.anthropic.com/en/release-notes for the GA announcement.

Methodology: when a number could not be cross-confirmed against an official Anthropic page on the verification date, it was omitted from this card rather than guessed.