Claude Opus 4.7: Full Spec Sheet (June 2026)

Opus 4.7 is the flagship of Anthropic's 'Claude 4' generation, succeeding Opus 4.5 (mid-2025) and Opus 4 (early 2025). It is the high end of a three-tier menu (Opus, Sonnet, Haiku) that Anthropic has maintained since Claude 3 in 2024 — Opus for hardest tasks at highest price, Sonnet for the production sweet spot, Haiku for cost-sensitive volume.

What makes Opus 4.7 different from GPT-5: stronger long-form writing voice, more disciplined instruction following on complex multi-step prompts, more aggressive prompt caching (90% off the cached read portion, vs OpenAI's also-90% but with different cache eligibility rules), native XML-tag structuring as the canonical prompt format, and extended thinking — Anthropic's name for explicit, configurable chain-of-thought reasoning that you can budget in tokens per call.

Where GPT-5 collapsed reasoning into a single `reasoning_effort` parameter, Anthropic exposes 'extended thinking' as a separate `thinking` block in the API: you set a `budget_tokens` (e.g. 5,000) and Opus 4.7 burns up to that many thinking tokens before producing the visible answer. Like reasoning tokens on GPT-5, thinking tokens bill at the output rate.

Standard rates: `cost = (input_tokens / 1M) × $15 + (output_tokens / 1M) × $75`. The representative 1,000-in / 500-out call: `0.001 × $15 + 0.0005 × $75 = $0.015 + $0.0375 = $0.0525`. About 5¢ per call — roughly 8× GPT-5 on the same tokens.

The price gap closes dramatically with prompt caching. Anthropic's caching is explicit and surgical: you mark a `cache_control: { type: 'ephemeral' }` block in the messages array, Anthropic stores that prefix server-side (5-minute or 1-hour TTL), and subsequent calls within the TTL read it at 10% of the input price ($1.50/M instead of $15/M). The first-time write costs $18.75/M (5-min) or $30/M (1-hour) — a 25% or 100% premium on the first write, amortized over every subsequent read.

Worked: a 5,000-token cached system prompt + tools, hit 100× in an hour with 1-hour TTL. First write: $0.15. Next 99 reads at $1.50/M × 0.005M = $0.0075 each = $0.7425 total. Total: $0.8925 vs $7.50 uncached. 88% off the input bill. The structural cost discipline that makes Opus economical: cache-anchor everything stable, put dynamic context last in user messages.

Apply Batch API on async workloads: 50% off both streams. A non-cached 1,000-in / 500-out batched call drops from $0.0525 to $0.02625. Stack caching + batching on the cached portion of a heavy workload and Opus 4.7 becomes economically viable for serious scale. Worked dollars across the Claude family: Claude API cost calculator.

Opus 4.7 ships with a 200,000-token context window. Unlike Claude Sonnet 4.6 (which supports an experimental 1M-token context via a beta header), Opus does not currently expose a 1M variant. For Anthropic users who need >200K, the path is Sonnet 4.6 with the 1M beta header, not Opus.

200K is enough to fit ~150,000 words (roughly 300 pages of single-spaced text), a moderate codebase chunk with metadata, or a multi-turn conversation history of ~50 turns at typical message lengths. It is smaller than GPT-5 (400K) and dramatically smaller than Gemini 2.5 Pro (1M).

Max output is 64,000 tokens per response — half of GPT-5's 128K ceiling. For long-form report generation, plan to chunk the output via multiple calls if you need >64K of generated content; Opus does not support arbitrary continuation within a single call.

Extended thinking lets you allocate a token budget for explicit chain-of-thought reasoning before Opus produces the visible answer. Configured via the `thinking` block in the API request: `thinking: { type: 'enabled', budget_tokens: 5000 }`. Opus uses up to 5,000 tokens of internal reasoning, then writes the answer. The thinking tokens are returned as a separate `thinking` content block (you can show, hide, or persist them) and bill at the output rate.

When to use extended thinking: complex code synthesis, math/proof tasks, multi-step planning with branching logic, legal/financial analysis where the reasoning chain is part of the deliverable, debugging tasks where you want to see where the model went wrong. Budget 3,000-10,000 tokens for hard tasks.

When NOT to use it: classification, extraction, summarization of structured input, anything mechanical. Adding thinking to a classification task adds cost with no quality improvement — Opus's instinct response on simple tasks is already calibrated.

Cost example: a 2,000-token problem with 5,000 thinking tokens + 800 visible output tokens bills `0.002 × $15 + 0.005 × $75 + 0.0008 × $75 = $0.03 + $0.375 + $0.06 = $0.465` per call. Use thinking deliberately; it is by far the highest-EV cost lever on Opus when applied to the right tasks.

Opus 4.7 supports the full Anthropic tool-use API: define tools as JSON Schema in the `tools` parameter, Opus picks one (or several in parallel) and returns the arguments in a `tool_use` content block. Parallel tool use is on by default.

Structured outputs on Anthropic are achieved via the tool-use mechanism: define a tool whose input schema matches your desired output schema, force the model to call it with `tool_choice: { type: 'tool', name: 'extract_data' }`, and parse the tool-call arguments. The JSON Schema enforcement is reliable as of Opus 4.7 — invalid outputs are extremely rare on well-formed schemas.

Anthropic's prompt convention emphasizes XML tags for structuring complex instructions: `<task>...</task>`, `<context>...</context>`, `<example>...</example>`, `<output_format>...</output_format>`. The model is trained to attend to these reliably, and Anthropic's docs show every advanced pattern using XML. Prompts written for GPT-5 (typically markdown-headed or just paragraph instructions) work on Opus, but Opus consistently performs better when the structure is XML-tagged.

**Pick Opus 4.7** when quality on hard tasks is the bottleneck and budget is not — agentic workflows that need long-horizon planning, complex code synthesis with multi-file context, legal/financial analysis with strict accuracy requirements, long-form writing where Anthropic's voice is preferred, deep research synthesis. The 5× price premium over Sonnet is justified when downstream cost-of-error dominates per-call cost.

**Pick Sonnet 4.6** ($3 / $15 per 1M) for the production sweet spot: structured-data pipelines, content generation, routine chat, summarization, classification with high quality requirements, anything where Anthropic's discipline matters but Opus is overkill. Most production teams who default to Claude live on Sonnet, not Opus.

**Pick GPT-5** instead of Opus 4.7 when: you need the larger 400K context window, you need native structured-output enforcement at the API level (slightly more mature than Anthropic's tool-use-as-structured-output pattern), or you're already deep in the OpenAI tooling ecosystem (Responses API, built-in tools, Assistants).

Cross-tier head-to-head: GPT-5 vs Claude Opus 4.7 covers the side-by-side on benchmarks, voice, and total cost of ownership.

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; docs.anthropic.com/en/docs/build-with-claude/prompt-caching for cache write/read mechanics and TTLs.

Anthropic publishes a changelog at docs.anthropic.com/en/release-notes that records pricing and model changes — more transparent than OpenAI's silent pricing-page updates. Subscribe or check monthly if your Opus bill is over $1,000/month.

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. If you find a discrepancy with the live page, treat the live page as canonical.