Claude Opus 4.7

Claude Opus 4.7 is still available in the API in August 2026. What it offers, how it differs from Opus 4.6 and Opus 5, and the one-line change to migrate.

Claude Opus 4.7 is still callable in the Claude API as of August 2026. The model ID is claude-opus-4-7, it costs $5 per million input tokens and $25 per million output, and Anthropic lists retirement as no sooner than 16 April 2027. It has been superseded twice over — first by Opus 4.8, then by Claude Opus 5, which is priced identically.

#What was Claude Opus 4.7?

Opus 4.7 arrived on 16 April 2026, between Opus 4.6 (5 February) and Opus 4.8 (28 May). Two changes shipped with it that outlived the model itself and now shape every current release.

The first was the new tokenizer. Anthropic's pricing documentation states that Claude 4.7 and later models use a newer tokenizer that "produces approximately 30% more tokens for the same text", with the exact increase depending on content and workload shape. Everything from 4.7 forward — Opus 4.8, Fable 5, Opus 5, Sonnet 5 — inherits it; anything at or before Sonnet 4.6 uses the previous one.

The second was the removal of sampling parameters. On Opus 4.7 and later, a non-default temperature, top_p or top_k returns a 400 error; the documented replacement is to omit them and steer through prompting. Opus 4.7 also introduced the task budgets beta (task-budgets-2026-03-13), and from 4.7 onward Opus-tier models stopped receiving the injected <budget:token_budget> and <system_warning> context-awareness tags that Sonnet and Haiku models still get.

#Opus 4.6 vs Opus 4.7 vs Opus 5

All three cost the same per token. The differences are in tokenizer, thinking behaviour, cache economics and which tools they can call.

The rows that matter most for cost are not the prices — they are the tool-use overhead and the cache minimum. Opus 4.7 carries the heaviest tool-use system prompt of any documented Opus model, against 286 and 406 tokens on Opus 5. For an agent making thousands of tool-bearing calls a day, that gap alone is real money. Reasoning in cost per task rather than cost per token is what the model comparison page is for.

#How do you migrate from Opus 4.7 to Opus 5?

Change the model string. Everything else is a short checklist.

msg = client.messages.create(
    model="claude-opus-5",   # was "claude-opus-4-7"
    max_tokens=8192,
    messages=[{"role": "user", "content": prompt}],
)

On Bedrock, anthropic.claude-opus-4-7 becomes anthropic.claude-opus-5; on Google Cloud, claude-opus-4-7 becomes claude-opus-5.

#What actually breaks

• Thinking is on by default on Opus 5. Prompts tuned against a non-thinking 4.7 will produce longer, more deliberate answers and more output tokens. Depth is set by effort: low, medium, high (default), xhigh, max. See the thinking and effort page.
• Turning thinking off is now conditional. thinking: {"type": "disabled"} is accepted only when effort is high or below; with xhigh or max it returns a 400.
• Sampling parameters stay gone. Anything already running on 4.7 has dropped temperature, top_p and top_k. Coming from 4.6 or earlier, this is the change that bites.
• Prompt-cache minimums drop from 2,048 tokens to 512. Prompts too short to cache before become cacheable, so short-system-prompt workloads can get cheaper just by adding a cache_control breakpoint. Multipliers are on the Claude API page.
• Remove carried-over verification instructions. Anthropic's migration guidance says lines like "include a final verification step" or "use a subagent to verify" cause over-verification on Opus 5.
• Token counts do not change. 4.7 and Opus 5 share a tokenizer, so a 4.7 → 5 move needs no re-count. A 4.6 → 5 move does.
• Rate limits move. Opus 4.5 through 4.8 share one combined bucket; Opus 5 has its own, so a phased rollout across both gains headroom.

#What Opus 4.7 cannot do in August 2026

Several capabilities that shipped after April 2026 skipped it.

• No computer or browser use. The GA computer_toolset_20260801 and browser_toolset_20260801 tools, released 19 August 2026, are documented for Fable 5, Mythos 5, Opus 5, Sonnet 5 and Opus 4.8 — not Opus 4.7.
• No fast mode. Fast mode is a research preview on Opus 5 and Opus 4.8 at $10 / $50 for up to 2.5× output speed. On Opus 4.7 it returns an error.
• The worst cache and tool economics of the Opus line, despite the identical sticker price.

Anthropic's statements about relative capability are qualitative, not numeric: the Opus 4.8 announcement says 4.8 is "approximately four times less likely than Opus 4.7 to overlook flaws in code", with the benchmark table published as an image rather than text. Precise benchmark scores for Opus 4.7 are not documented in current sources, and this site does not repeat third-party figures that cannot be traced to a primary Anthropic page. Where Claude wins and loses against rival vendors is on the Claude vs ChatGPT comparison.

#When staying on Opus 4.7 is defensible

"Always upgrade" is bad engineering advice, and there is a year of runway. Three cases hold up.

#You have a validated prompt on a pinned snapshot

Dateless model IDs from the 4.6 generation onward are pinned snapshots, not evergreen pointers — Anthropic does not update weights under an existing ID. If a production prompt cleared a regulated sign-off or sits inside a contractual SLA, claude-opus-4-7 is a fixed target until April 2027, and moving means re-validating. One caveat Anthropic states explicitly: weights are fixed per ID, but serving infrastructure — router, safety classifiers, sampling logic — can change and produce minor observable differences. Pinned is not frozen.

#Your evaluation baseline is expensive to rebuild

If your regression suite was scored against Opus 4.7 outputs, and especially if human graders set the pass bar, switching generations invalidates the baseline. Thinking-on-by-default alone moves pass rates for reasons unrelated to quality. Keep 4.7 serving production, run Opus 5 in shadow against the same suite, and cut over on evidence — the two sit on separate rate-limit buckets, so shadow traffic costs you no headroom.

#Budget predictability during a spend freeze

Opus 4.7 and Opus 5 cost the same per token, so this argument is weaker here than on the Sonnet 4.6 page — but not empty. Thinking on by default means more output tokens per task unless you explicitly lower effort, and output costs five times input. Inside a fixed cap, staying put keeps the bill flat until you have measured the new output profile. Rates are on the pricing page.

What is not defensible: staying because migration looks hard. Claude Code ships a /claude-api migrate skill that applies model ID swaps and breaking parameter changes across a codebase — see Claude for developers. The full Opus history, including which IDs now return errors, is on the Opus version archive; the current lineup is indexed on the Claude models page.

#Frequently asked questions