A request path being quietly redirected through a gateway toward a veiled model, illustrating the Claude Opus 5.2 routing reports

model-release

Claude Opus 5.2 Release Date: Rumor vs What Anthropic Says

EvoLink Team

Product Team

September 17, 2026

Updated on September 18, 2026

17 min read

Claude Opus 5.2 has not been released or announced. Since September 14, 2026, several X accounts have reported that Claude Code is quietly routing some Opus 5 requests to a build they call "Opus 5.2". In the Anthropic sources we checked on September 17–18, 2026 (model catalog, release notes, pricing page, newsroom, SDK model lists, and the Claude Code 2.1.276 binary) we found no Opus 5.2 entry and no claude-opus-5-2 identifier. The newest Opus release is Claude Opus 5 from July 24, 2026.

If you run Opus 5 in production, this page gives you the current answer, the evidence behind it with links you can re-check, the signals that would change it, and the separate stages between an announcement and a request you can bill. Access, pricing, and model-ID status live on the Claude Opus 5.2 API page; how to decide whether a point release deserves your Opus 5 traffic is in Claude Opus 5.2 vs Claude Opus 5.

The Direct Answer: What Is Confirmed Today?

Question	Answer (checked September 17–18, 2026)	Evidence level
Has Anthropic announced Claude Opus 5.2?	No entry found in the newsroom, model catalog, release notes, pricing page, or Claude Code changelog	Not found in the official sources checked
Is there a `claude-opus-5-2` model ID?	Not found in Anthropic docs, the Python and TypeScript SDK model lists, the three cloud catalogs, or the Claude Code 2.1.276 build	Not found in the official sources checked
Is Claude Code routing Opus 5 to a new build?	Unverified. Several X users report it; Anthropic has not commented; no post includes an identifier or request record	Community reported
Did a Foundry file confirm the model?	No. The file is a community registry entry whose values are identical to Opus 5	Third-party source, traced and linked below
Does 5.2 add xhigh or max effort?	Those levels are already documented for Opus 5	Official (effort documentation)
Release date	Not published. Guesses posted September 14–15 ranged from September 17–18 to "next week"; none is an announcement	Community reported
Pricing, context, limits	Not published	Unknown
EvoLink route	No verified route, SKU, price, or date	EvoLink status, separate from Anthropic

Every positive fact in this table is about Opus 5. Everything attributed to Opus 5.2 is either "not found where we looked" or a community report. "Not found" is a dated observation about specific sources, not a claim about Anthropic's roadmap.

What the Claude Opus 5.2 Routing Reports Actually Say

The rumor has a traceable starting point. The posts below were read directly on X on September 17, 2026.

Date (UTC)	Account	What was claimed
September 14, 12:15	@notjazii	Compared the same prompt "on xhigh" in Devin and Claude Code, found the Claude Code result cleaner, and concluded "opus 5 is getting routed to opus 5.2"
September 14, 14:00	@notjazii	Called the routing "confirmed" after more of the same tests: "way faster", "really clean output", "not lazy"
September 14, 17:41	@synthwavedd	"likely being greyscale tested on Claude Code under the 'Opus 5' name"; says web requests still reach the older model. The most-shared post of the cycle
September 15, 16:04	@pankajkumar_dev	"overthinks less", "much faster", "not lazy anymore"; guesses "next week"

Likes and views do not make a model claim more likely to be true, so we do not rank the posts by them. What matters is what they contain:

No identifier, request record, or Anthropic statement. The evidence is output quality compared by eye, mostly by one account on one day.
The "confirmation" is the first account repeating its own test. No independent source has produced a different kind of evidence.
Scope is disputed. One post says the web app still serves the older model. Forum threads from the same days include users who say the change reached the web app but not Claude Code, users on Pro plans who say they have it, and at least one Chinese-language forum user who reported having it one evening and losing it again afterwards.

A "this Thursday or Friday" estimate is also circulating. We could only find it relayed by third-party blogs, not in a post we could open, so treat it as a second-hand prediction. It pointed at September 17–18, 2026. When we checked the sources on September 18, no announcement had appeared.

What Anthropic's own documentation says about this kind of change

The most useful official text for reading these reports is not about Opus 5.2 at all. Anthropic's model IDs and versioning page states two things:

A model ID is a pinned snapshot: "Anthropic does not update the weights or configuration of an existing model ID. When an updated version is available, it ships under a new model ID."
Behavior can still move: "Model weights are fixed for a given ID, but the serving infrastructure around the model can change over time," including "the request router, safety classifiers, and sampling logic," and when behavior differs on a stable ID, "an infrastructure update is the most likely cause."

That guarantee is written for API model IDs. It does not describe how the Claude apps or Claude Code choose a model behind a product label, so it neither confirms nor rules out an unlabeled test inside Claude Code. It does mean that "the same name suddenly feels faster" has a documented, ordinary explanation that does not require a new model, and that an API request for claude-opus-5 is documented to reach the Opus 5 snapshot.

The "Foundry Slug", Traced

The most repeated piece of hard-looking evidence is a claude-opus-5-2 slug, sometimes described as a claude-opus-5-2.yaml file listing five reasoning-effort tiers, said to have appeared in Microsoft Foundry. We traced it, and you can re-check each step.

What we observed

The only claude-opus-5-2.yaml files we could find on GitHub are in a community-maintained model registry that describes itself as "metadata only", not in a Microsoft or Anthropic repository: the microsoft-foundry copy and the azure-ai-foundry copy. Both commits were authored by an automated bot, on September 4 and August 31, 2026.
The file's values are Opus 5's documented values: $5 input and $25 output per million tokens, a 1M context window, 128K max output, and effort levels low, medium, high, xhigh, and max. Its own sources field links to the Opus 5 catalog page.
The same bot created -2 files for exactly four Claude models in the same weeks: Opus 5, Opus 4.8, Sonnet 5, and Haiku 4.5. Microsoft's Foundry documentation offers Claude models in two versions, version 1 hosted on Anthropic infrastructure and version 2 hosted on Azure, and lists those same four models as Azure-hosted.
Microsoft's Claude-in-Foundry model table, updated September 11, 2026, lists claude-opus-5 and no claude-opus-5-2. The Foundry catalog's Anthropic publisher page lists 14 models, none of them 5.2.
ai.azure.com/catalog/models/claude-opus-5-2 returns a page with the slug in its title, and so do claude-opus-9-9 and a made-up name. The catalog is a single-page app that echoes any slug, so an opening URL is not evidence.

What this does and does not establish

The registry bot's source code is not public, so we have not proven why it named the file -2. The observation is that the suffix lines up one-to-one with Foundry's hosting-version numbering and that every value in the file is an Opus 5 value. That is enough to say the file is not evidence of a new model. It is not enough to say how the bot works, and we are not saying anyone fabricated it.

The five effort levels are the same story. Anthropic's effort documentation already lists low, medium, high, xhigh, and max for Claude Opus 5; xhigh arrived with Opus 4.7 in April 2026.

Claude Opus 5.2 evidence checking: unconnected labels are separated from source documents, request records, and billing checks; a conceptual illustration, not model identity proof

Why the Name Might Not Survive: The Opus 5.1 Precedent

In late August 2026 the same pattern ran under a different name. Widely shared posts described early-access strings and A/B traffic for an "Opus 5.1". On September 1, Anthropic released Claude Fable 5.1 and Mythos 5.1. Nothing was released under the name Opus 5.1, and the Opus line still ends at Opus 5. The separate next-generation rumor is tracked in the Claude Opus 6 release watch.

Two consequences for anyone planning around "5.2":

The number is a community guess. Several posts openly ask why it is 5.2 and not 5.1. Nobody outside Anthropic knows what the build is called, or whether the faster behavior is a new model at all rather than an infrastructure, system-prompt, or harness change.
Wait for the published name and request ID. Anthropic's current format is claude-{name}-{major}[-{minor}], which is why people expect claude-opus-5-2. A naming convention explains what a string would mean; it does not commit Anthropic to a product, a name, or an ID. Do not put a guessed string in configuration.

What Would Confirm a Claude Opus 5.2 Release

Any one of these is enough to update the status, and they do not have to appear together:

An Anthropic newsroom post, or a model page with a "Released" date, naming a new Opus.
A new Opus entry in the platform release notes, the pricing page, or the model-deprecations table.
A new Opus ID in the official Python or TypeScript SDK model lists or in a Claude Code release.
A new Opus row in Microsoft's Claude-in-Foundry table, the Amazon Bedrock model list, or Google Cloud's Claude model page. Record each channel separately; do not wait for all of them.
An EvoLink route verification: an approved model ID, a SKU, a successful authenticated request, and usage and billing that reconcile.

We re-check these sources when a new signal is reported and show the date of the last check at the bottom of this page. The URL and publication date stay fixed; the "updated" date changes only when the evidence does.

A Claude Opus 5.2 Release Would Not Mean API Availability

An announcement settles the name and the launch scope. It does not mean EvoLink can call the model. Three stages are verified separately.

Stage 1: Anthropic announcement. A newsroom post, a model page, and a release-notes entry. Each Opus launch from 4.5 through 5 was announced as available on the Claude API and the major cloud platforms the same day, but that is history, not a rule; treat each channel's listing as its own fact.

Stage 2: Upstream API behavior. The model ID, pricing row, effort and thinking defaults, context and output limits, and the what's-new page. Past Opus releases changed integrations in three different ways, and it helps to keep them apart:

Kind of change	Documented examples
Documented breaking change	Opus 5: thinking on by default; thinking can be disabled only at effort high or below (400 otherwise). Opus 4.7: non-default `temperature`, `top_p`, `top_k` return a 400
Cost or behavior change	Opus 4.7: new tokenizer, so the same text counts as more tokens. Opus 5: thinking tokens are billed as output tokens, so a workload that ran without thinking produces more output tokens at the same rates
New capability	Opus 4.8: fast mode (also offered on Opus 5). Opus 5: 512-token cache minimum, mid-conversation tool changes (beta)

Only the first row breaks a request that used to succeed. The second row changes your bill without an error. A future point release could do any of the three, or none.

Stage 3: EvoLink route. An approved model ID, a SKU and price, a successful authenticated request, and usage and billing records that reconcile. An upstream listing tells EvoLink what to verify; it does not verify it.

Three separate Claude Opus 5.2 access checks: provider announcement, documented API behavior, and EvoLink route verification, with the final gate closed to represent unverified access

The Claude Opus 5.2 API page tracks each stage with a dated status and keeps saying "no verified route" until Stage 3 passes.

How to Check Whether You Are Being Routed

The community's favourite test is to ask Claude Code, without web search, who "Tibo the reset guy" is; a correct answer is taken as proof of the newer model. Treat it as a curiosity:

It reads generated text. Users report different answers to the same question on the same machine across Claude Code versions, which points at prompt or harness differences as much as at a model swap.
Anthropic documents Opus 5's reliable knowledge cutoff as May 2026, and the supposed new build has no documented cutoff to compare against.
A name is not a guarantee of what serves a request: a user-created bot on Poe called "Claude-Opus-5.2" describes itself as powered by an OpenAI model, and at least one API reseller shows a "Claude Opus 5.2" model page with a copy-model-ID button while its own status text says the official API is not announced.

For an API route, the checks worth running are consistency checks. None of them is independent proof of the underlying weights, because every field is supplied by the server or a proxy in front of it, but together they make a substitution traceable:

The requested ID matches a documented ID or channel mapping (Bedrock and Google Cloud use their own ID formats).
The returned model field and request ID are logged for every call, and they agree with the documented mapping.
Usage and billing reconcile against the published price for that ID.
The same task set at the same effort level behaves consistently over several days, not in one session.

What Developers Can Prepare Now

The work is the same whether the build ships as 5.2, as 5.1, or not at all, and the comparison guide covers it in detail. In short:

Keep the candidate model ID in configuration and leave it empty. Do not pre-fill claude-opus-5-2.
Baseline Claude Opus 5 at the effort levels you actually use, including how often long tasks needed a "continue". You cannot measure an improvement you have not baselined.
Save the traces where Opus 5 disappointed you. "Lazy", over-thinking, and hard to steer are the claims a successor has to beat.
Test your rollback route, do not assume it. Anthropic lists claude-opus-5 as active with retirement "not sooner than July 24, 2027" on Anthropic-operated platforms (the Claude API, Claude Platform on AWS, and Microsoft Foundry); Amazon Bedrock and Google Cloud set their own schedules. That gives you an evaluation window. It does not guarantee the capacity, permissions, or status of the specific EvoLink route you would roll back to, so verify that route before you need it.

Track Claude Opus 5.2 API availability on EvoLink Compare current Claude models

FAQ

What is Claude Opus 5.2?

It is a community name for a build that some X users believe Claude Code has been serving under the "Opus 5" label since about September 14, 2026. It is not an Anthropic product name: we found no model page, model ID, price, or announcement for it in the official sources checked on September 17–18, 2026.

Is Claude Opus 5.2 released?

No announcement exists in the Anthropic sources we checked on September 17–18, 2026: the model catalog, release notes, pricing page, newsroom, and SDK model lists. The newest Opus is Claude Opus 5, released July 24, 2026.

Is Claude Code really routing Opus 5 requests to Opus 5.2?

It is unverified. Several X users reported faster, less lazy Opus 5 behavior in Claude Code from September 14 and attributed it to a hidden build. Anthropic has not commented, no post includes an identifier or request record, and reports disagree on which apps and plans are affected. Anthropic's versioning docs note that serving infrastructure can change behavior on a stable model ID.

Why does Opus 5 feel faster in Claude Code?

Nobody can tell from the outside. A hidden model test is one explanation. Anthropic's versioning documentation offers another: serving infrastructure such as the request router and sampling logic can change behavior on a stable model ID. Claude Code updates and system-prompt changes are a third. Without a request record you cannot tell them apart.

Which plans do the reports mention: Pro or Max, web or Claude Code?

Reports conflict. One widely shared post says web requests still reach the older model; forum users report the change on the web app but not in Claude Code, on Pro plans, and in third-party agents, and at least one says it disappeared again. Anthropic has documented no test, so no plan or app can be said to include it.

Why is it called 5.2 and not 5.1?

Nobody outside Anthropic knows. The name comes from community posts and a registry file. In August a rumored "Opus 5.1" was never released under that name; Claude Fable 5.1 was. Wait for the name and request ID Anthropic publishes.

Did Microsoft Foundry list claude-opus-5-2?

We did not find it in Microsoft's Claude-in-Foundry table (updated September 11, 2026) or on the Foundry catalog's Anthropic publisher page. The claude-opus-5-2.yaml file circulating online is a bot-generated entry in a community registry whose values equal Opus 5's, and its -2 suffix lines up with Foundry's "version 2" label for Azure-hosted deployments.

Does the "Tibo" question prove I have the new model?

No. It reads generated text, and users report different answers across Claude Code versions on the same machine. For API traffic, check that the documented ID mapping, the returned model field, and billing agree; even those are consistency checks, not independent proof of the weights.

When will Claude Opus 5.2 be released?

No date is published. An X post from September 15 guessed "next week", and a "this Thursday or Friday" (September 17–18) estimate circulated second-hand through third-party blogs. When we checked on September 18, no announcement had appeared. On September 17, prediction markets priced a new Opus by September 30 at roughly 86% and a release named exactly "Opus 5.2" in September at 32%. These are predictions, not announcements.

Will Opus 5.2 cost the same as Opus 5?

Unknown. Every Opus release from 4.5 through 5 listed $5 input and $25 output per million tokens, which is a reasonable input for a budget model and is not an Opus 5.2 price. Use the price and availability Anthropic actually publishes, if and when it does.

Can I use Claude Opus 5.2 through EvoLink now?

No. EvoLink has no Opus 5.2 route, SKU, price, or integration date. The Claude Opus 5.2 API page tracks the status and offers a launch alert; the current Opus route is Claude Opus 5.

Sources

Official sources last checked September 18, 2026; X posts read September 17, 2026. Anthropic and Microsoft facts come from their documentation; community reports are labeled as such; EvoLink route status is separate and unverified.

All Posts

#Claude Opus 5.2 #Claude Opus 5 #Anthropic #Release Date #Release Watch