Claude Opus 4.7 Is Still Active. What Changed, and When Should You Migrate?

Last updated: August 6, 2026

Claude Opus 4.7 is an Anthropic language model released on 16 April 2026. As of 6 August 2026 Anthropic lists it as Active, with no deprecation date and a tentative retirement date no earlier than 16 April 2027. Anthropic has since released Opus 4.8 and Opus 5. All three carry the same base API rate.

Verdict: An existing standard-speed Opus 4.7 deployment faces no first-party retirement deadline today. Active is a lifecycle status, not a compatibility warranty, so the migration question should be settled by dependency audit rather than by model age or benchmark headline.

Sources: Anthropic, 16 April 2026 · Anthropic, Model deprecations, read 6 August 2026 · Anthropic, Pricing, read 6 August 2026

Sources: Anthropic, Release notes, entries 16 April, 12 May, 25 June and 24 July 2026 · Anthropic, 16 April 2026 · Anthropic, Pricing · Artificial Analysis, 15 June 2026

Sources: Anthropic, Migration guide · Anthropic, Choosing a model · Anthropic, Model deprecations

What Active states, and what it does not

Anthropic publishes four lifecycle states. Active is defined as fully supported and recommended for use, followed by Legacy, Deprecated, and Retired. Opus 4.7 sits in the first state with no deprecation date assigned. Anthropic also states that customers with active deployments receive at least 60 days of notice before a publicly released model is retired.

Those sentences are worth reading for what they do not extend to. The lifecycle page does not state that a preview feature will remain available for the life of the model. It does not state that every request field accepted on the launch date will continue to be accepted. It does not state that partner platforms will use the same dates, and in fact says the opposite: the published schedule applies to Anthropic-operated surfaces, while Amazon Bedrock and Google Cloud set their own.

There is one place where Anthropic has closed a related ambiguity. Model identifiers are documented as pinned snapshots rather than pointers that quietly advance to a newer model, including the dateless identifiers introduced with the 4.6 generation. A team that pins claude-opus-4-7 is calling a fixed model. The precise buyer statement is therefore narrower than either optimism or suspicion would suggest. The model snapshot is fixed and supported. The commercial service around it, meaning preview features, accepted request configurations, platform schedules, pricing modifiers, and the external measurements people quote, is governed separately.

The platform point is worth making concrete, because the two lifecycle systems do not share a vocabulary. Anthropic publishes four states. Amazon Bedrock publishes three: Active, Legacy, and End-of-Life. Bedrock’s commitment is shaped differently too. Where Anthropic gives a tentative retirement floor plus at least 60 days of notice, AWS states that once a model launches on Bedrock it remains available for at least 12 months before its EOL date. Bedrock also exposes the state programmatically, in a lifecycle field returned by its foundation-model API calls, which Anthropic’s own surface does not offer. A team running the same model on both is holding two different clocks, and only one of them can be queried from code.

Sources: Anthropic, Model deprecations, read 6 August 2026 · Anthropic, Models overview · AWS, Amazon Bedrock model lifecycle

Fast mode and the failure-semantics gap

Anthropic added fast mode support for Opus 4.7 on 12 May 2026 as a research preview behind a beta header, deprecated it on 25 June, and removed it on 24 July. Twenty-nine days of notice, published in the release notes rather than in the model lifecycle table.

The removal is defensible on its own terms. The feature was labelled a research preview from its first entry. Preview features carry preview risk, and nothing in the lifecycle definition covers them.

The engineering finding sits one level down, in how the removal was implemented. Anthropic’s release note draws the contrast itself: requests to Opus 4.7 with the fast speed setting now return an error and, unlike Opus 4.6, do not fall back to standard speed. The same option, withdrawn from two adjacent models, produces two different outcomes. On 4.6 a stale request degrades silently and bills at the standard rate. On 4.7 it fails.

For an integration this matters more than the removal date. A routing or retry layer written against the 4.6 behavior handles a fallback that no longer occurs, which means the code path most likely to still carry a fast speed setting is also the code path least likely to be watching for an error. The documents establish the difference in API behavior. What any given production system logged or surfaced is not something FSR can determine from documents.

Sources: Anthropic, Release notes, entries 12 May, 25 June and 24 July 2026 · Anthropic, Pricing, fast mode availability

Choosing between 4.7, 4.8 and 5

All three models carry the same base rate. The choice is therefore not a price comparison but a dependency comparison.

Staying on Opus 4.7 is defensible when the deployment uses standard speed, still passes your own evaluations, sits on a platform whose retirement schedule you have checked, and needs nothing that only a newer model provides. Anthropic’s first-party lifecycle table imposes no deadline on that position today, which is not the same as evidence that staying is optimal.

Opus 4.8 is the smaller move. It launched with the same base rate, the same 1M context window and 128k output limit, the same tool and platform set as 4.7, and it retains fast mode as a research preview. Smaller is a hypothesis about migration cost, not an FSR measurement of it.

Opus 5 is the larger re-baseline. Anthropic prices it identically but documents behavioral and entitlement differences: adaptive thinking is on by default, disabling thinking at xhigh or max effort returns a 400 enforced on every individual request, and web fetch and Priority Tier are not available. Anthropic describes out-of-the-box performance as strong while still instructing developers to run a fresh effort sweep and re-evaluate cost and latency rather than carrying settings across.

One property of Opus 5 has no equivalent on 4.7 and deserves its own line in a dependency audit. Opus 5 ships with cybersecurity safety classifiers, and Anthropic documents that cyber-category refusals can fall back to Opus 4.8. A refusal returns as a successful HTTP 200 with a refusal stop reason rather than an error. The billing rule is documented and specific: input tokens are not billed when a request is refused before any output is generated, but when a classifier fires mid-stream, the input and the already-streamed output are billed and the partial output is discarded. A team migrating from 4.7 acquires a failure mode that bills, returns 200, and may be answered by a different model than the one requested.

Sources: Anthropic, Migration guide, read 6 August 2026 · Anthropic, What’s new in Claude Opus 4.8 · Anthropic, Pricing

Why the base rate does not settle cost

Opus 4.7, Opus 4.8, and Opus 5 all list at $5 per million input tokens and $25 per million output tokens. That figure answers a procurement question and not a budgeting one.

Six documented modifiers sit between the rate card and an invoice, and they move in both directions. Cache writes cost more than fresh input, at $6.25 per million for the five-minute tier and $10 for the one-hour tier, while cache hits cost $0.50, one tenth of standard input. Batch processing halves both sides, to $2.50 and $12.50. US-only inference adds a 1.1 multiplier on qualifying first-party surfaces. Tool definitions add a fixed system-prompt overhead of 675 or 804 tokens depending on configuration, with the bash tool adding 325 more and the text editor tool 700. Automatic high-resolution image input can consume roughly three times the image tokens of earlier models. Effort level changes how many output tokens a task produces.

The tokenizer is the modifier most often quoted and least often bounded. Anthropic’s launch post gave a range of roughly 1.0 to 1.35 times as many tokens for the same input. Current pricing documentation compresses that to approximately 30% more tokens for the same text. Neither page publishes a distribution by content type, so a buyer cannot locate their own workload inside the range.

An inflated input meter suggests a higher bill, and one measurement points the other way. Artificial Analysis reported that running its then-current Intelligence Index on Opus 4.7 used 102 million output tokens against Opus 4.6’s 157 million, at a total cost of roughly $4,406 against roughly $4,970. On that suite, at max effort, under the April index and its cost methodology, the newer model finished cheaper despite the tokenizer.

That result is a data point about one benchmark, not a forecast for a production workload. What it establishes is the shape of the question. Task cost is a function of input tokens, cache behavior, thinking and answer tokens, tool overhead, regional modifiers, retries, and failure rate. FSR has measured none of those on production traffic, and neither Anthropic nor Artificial Analysis publishes cost per completed task for an application resembling yours.

Sources: Anthropic, Pricing, read 6 August 2026 · Anthropic, 16 April 2026 · Artificial Analysis, 17 April 2026

Reading a benchmark score correctly

Sources: Artificial Analysis, 17 April 2026 · Artificial Analysis, 15 June 2026 · Artificial Analysis, Claude Opus 4.7, read 6 August 2026

Questions answered

Sources: Anthropic, Model deprecations · Anthropic, Release notes · Anthropic, Migration guide · AWS, Amazon Bedrock model lifecycle · Artificial Analysis, 15 June 2026

Methodology

This is a Tier C briefing. FSR did not run Claude Opus 4.7 and did not measure its availability, reliability, latency, output quality, or production cost. Every operational statement traces to vendor documentation, a dated release note, or a named third-party measurement, read on 4 and 6 August 2026.

Facts are separated by the date they became true and by who made them true. Changes Anthropic shipped at launch are not presented as changes that happened to an Active model afterwards. Changes made by a measurement organization are not presented as changes to the product.

The Opus 4.7 system card was opened and read directly for this briefing, as were Anthropic’s migration guide and the Amazon Bedrock model lifecycle page. Two limits remain. The minimum cacheable prompt length for Opus 4.7 is not published; documentation establishes only that it sits above the 1,024-token minimum on Opus 4.8. And Anthropic’s pricing and lifecycle pages carry no visible last-updated date, so both should be reopened before a production decision.

Sources: Anthropic, 16 April 2026 · Anthropic, Release notes · Anthropic, Model deprecations · Anthropic, Pricing · Anthropic, Migration guide · Anthropic, Models overview · Anthropic, Choosing a model · Anthropic, What’s new in Claude Opus 4.8 · Anthropic, Claude Opus 4.7 System Card, 16 April 2026 · Artificial Analysis, model page · Artificial Analysis, 17 April 2026 · Artificial Analysis, 15 June 2026 · AWS, 16 April 2026 · AWS, Amazon Bedrock model lifecycle

Verdict

Anthropic still lists Opus 4.7 as Active, with a tentative retirement no earlier than 16 April 2027 and at least 60 days of notice before that happens. On the documents, a standard-speed first-party deployment faces no deadline this quarter.

What the documents also show is that four things sit under a single lifecycle label and are governed separately. The model snapshot is pinned. The preview features attached to it are not. The request configurations the API accepts changed at launch and can change again. The retirement date depends on which platform you call. And the benchmark score people quote depends on which version of an external index produced it.

The useful discipline is to stop asking one question and start asking four. Is the model still served, is my request still accepted, does my platform use the same schedule, and does my cost per completed task still hold. The first three have published answers. The fourth has to be measured on your own traffic, and nothing in this briefing substitutes for that.