Claude Opus 5.5: Skipping 5.2, Cutting Prices, Aiming at GPT-6
Opus 5.5 is not released, but every signal points to the week of September 21, 2026: stealth-testing under `claude-wafer-eap`, a rumored $4/$20 price, and prediction markets at 90%. Here is what is real, what is not, and which Claude models and prices you can use right away.
Published 2026-09-21
Anthropic's next flagship is Opus 5.5, and the evidence points to it landing in the week of September 21, 2026. It is not released yet — there is no official model ID, price, model card or date, and the strings Opus 5.5, claude-opus-5-5, claude-wafer-eap and Fable 5.2 appear zero times in Anthropic's public documentation. But this is not the usual thin rumour: on September 20, 2026 a leaker with a live track record said Opus 5.5 is in stealth testing under the codename claude-wafer-eap with a Tuesday target date, two independent accounts corroborated the pricing, and the prediction market that settles on Anthropic's own announcements priced a release at 79.5% by September 24 and 90% by September 30.
Three claims define the story, and they are worth separating before the launch noise arrives:
- The version is 5.5, not 5.2. Everyone expected 5.2 three weeks ago. The next public Claude is now expected to carry a much bigger number, which is a statement about how far Anthropic jumped rather than a normal increment.
- The price is reportedly going down, not up. $4 input and $20 output per million tokens, against the official $5/$25 for Opus 5 — with cache reads at $0.20, the number that actually changes agent economics.
- The pricing source disclaims itself. The graphic carrying $4/$20 is stamped Unofficial concept · Unconfirmed details. Read it as a credible directional signal, not a quote.
You do not have to wait for any of it. Claude Opus 5 starts at $1.48467 per million input tokens on APIMaster — about 70% below Anthropic's official $5/M — across 11 live routes on the Opus 5 marketplace card (checked September 21, 2026; other channels for the same model run up to list price, so the card is authoritative). That is roughly 63% cheaper than the $4/M number being floated for 5.5, which is the part of this story that is easy to miss. Sonnet 5, Sonnet 4.6, Opus 4.7, Opus 4.8 and Haiku 4.5 start at about 91% below official. Create an account, generate a key in the console, and one OpenAI-compatible endpoint covers Claude, GPT, Gemini, DeepSeek and Kimi.
GPT Trial(GPT-6 Astra, GPT-5.6, GPT-5.5, and GPT Image 2 available)
Sign up and claim a $20 GPT trial
What is confirmed and what is not
| Item | Status as of September 21, 2026 |
|---|---|
| Official Anthropic announcement of Opus 5.5 | None |
claude-opus-5-5 as a published model ID |
Not published |
| Official price for Opus 5.5 | Not published |
| Official release date | Not published |
| Context window, output limit, benchmarks | Not published |
Stealth-test report naming Opus 5.5 and claude-wafer-eap |
Posted September 20, 13:43 UTC by a tracker with two correct calls in the same week |
| Pricing of $4 in / $20 out / $0.20 cache read / $5 cache write | Posted September 20, 15:30 UTC, on a graphic labelled Unofficial concept · Unconfirmed details |
| That the next model is 5.5 rather than 5.2 | Community consensus since September 20 — no official basis either way |
| Release "planned for Tuesday, September 22", possibly September 21 | Leak claim; prediction markets agree on the week, not the day |
| Newest Claude model you can call | Claude Fable 5.1 (claude-fable-5-1), released September 1, 2026 |
The bottom half of that table is why this page exists. It is also the half that most of the coverage is treating as the top half. A codename, a partner-program entry and a pricing graphic are claims about a pipeline — strong ones, but still claims.
What the September 20 leak actually says
We read the primary posts rather than the coverage of them. The sequence, in UTC:
- 13:43:50 —
@lyraxana: "Anthropic is currently stealth testing Opus 5.5 (claude-opus-5-5) under the codenameclaude-wafer-eapwhich is planned to be released on Tuesday." - 15:30:41 —
@synaptized: "Opus 5.5 Pricing: Input per Mtok: 4$ / Output per Mtok: 20$ / Cache Read: 0.2$ / Cache Write: 5$" - 15:35:40 —
@lyraxana, adding detail: "Anthropic added "imminent" note and partners only got access for 24h / Dated for Tuesday, actual release could be Monday (!!) as well. Either way, it's the upcoming week / Thinking is ALWAYS adaptive (no "off" mode) / Forced tool use is retired / Pricing below" - 16:15:56 —
@chetaslua: "Opus 5.5 < previously we thought it will be 5.2> ... if you guys launch video for opus 5.5 i have one output it will blows everyone mind"
Three details in that sequence carry more weight than the headline.
The window is described as movable. "Dated for Tuesday, actual release could be Monday" — September 22, with September 21 floated as an earlier possibility. That is consistent with an "imminent" flag in a partner console rather than a scheduled launch event, which is what makes the earlier-than-expected case credible. Note the ordering: the leak itself puts the release no later than September 22, 2026.
Partner access is described as a 24-hour window. That is the shape of a controlled early-access exercise, not a rollout. It also means that if you see Opus 5.5 in a gateway while official channels still say nothing, you are looking at the program rather than the product.
The pricing graphic disclaims itself. The card that introduced $4/$20 is stamped Unofficial concept · Unconfirmed details. Almost every repost of those numbers leaves that line out. Hold both facts at once: the direction is probably right, the figure is not a price.
On why the source deserves attention rather than dismissal: the same account posted on September 17 that a new Gemini 4 Pro checkpoint was in testing, and on September 18 that "Nano Banana 2.5 (spicy-mayo) has been deployed... Partners got access through Vertex under the model name nano-banana-2.5." Two pre-announcement calls in the same week is a real signal about a source. It is still evidence about the source — codenames leak and then change, and a leak is not a launch.
Why 5.2 disappeared
For most of September the expected next Claude was 5.2. Our Claude 5.2 tracker documented routing reports naming Opus 5.2 from around September 14, a claude-opus-5-2 slug appearing in a partner catalogue, and a Reuters report that Anthropic was weighing an additional release without naming a model.
The September 20 posts reframe those reports rather than contradict them: @chetaslua announced it plainly as "previously we thought it will be 5.2." The reading that fits everything on the public record is that a 5.2 checkpoint was tested and that the next public name carries a bigger number — a jump large enough to be deliberate, not a routine increment. That reading is inference. Anthropic has never commented on 5.2 either.
The useful takeaway is not which number is right. It is that the community had the name wrong three weeks ago, and that the same routing-label logic producing "5.5" now is the logic that produced "5.2" then. Track the release, not the rumour.
The leaked price, and why the cache number matters more
Pricing is the part of a launch that changes your bill immediately, so it is worth reading carefully.
| Official Opus 5 | Rumored Opus 5.5 | Direction | |
|---|---|---|---|
| Input per 1M | $5 | $4 | ↓ 20% |
| Output per 1M | $25 | $20 | ↓ 20% |
| Cache read | — | $0.20 | new low for the tier |
| Cache write | — | $5 | — |
| Context window | 200K-class | 872K reported (down from a reported 936K) | ↓ |
A 20% input cut is a modest, plausible move — it matches the pattern Fable 5.1 set, where the headline price stayed flat and the cost improvement arrived through cache reads. The $0.20 cache read is the number that would change architecture decisions. For a long-running agent loop that re-sends large context on every step, cache reads dominate the bill, and an order-of-magnitude cut there is the difference between a demo and a product. If that number survives contact with the launch page, it matters more than the headline figures everyone is arguing about.
For reference, the live numbers on the first day of that expected release window — every figure pulled from APIMaster's marketplace data for the model ID, filtered to channels that are actually selling, checked September 21, 2026:
| Model | Official (input / output per 1M) | Live routes | APIMaster from | Saving | Card |
|---|---|---|---|---|---|
| Claude Fable 5.1 | $10 / $50 | 6 | $2.8944/M input | up to ~71% | Fable 5.1 |
| Claude Fable 5 | $10 / $50 | 4 | $6.5424/M input | up to ~35% | Fable 5 |
| Claude Opus 5 | $5 / $25 | 11 | $1.4847/M input | up to ~70% | Opus 5 |
| Claude Sonnet 5 | $2 / $10 | 12 | $0.1875/M input | up to ~91% | Sonnet 5 |
| Claude Sonnet 4.6 | $3 / $15 | 11 | $0.2813/M input | up to ~91% | Sonnet 4.6 |
| Claude Opus 4.7 | $5 / $25 | 13 | $0.45/M input | up to ~91% | Opus 4.7 |
| Claude Opus 4.8 | $5 / $25 | 13 | $0.45/M input | up to ~91% | Opus 4.8 |
| Claude Haiku 4.5 | $1 / $5 | 7 | $0.09/M input | up to ~91% | Haiku 4.5 |
| Claude Opus 5.5 | Not published | 0 | Reserved — this row gets a number when Anthropic publishes one | — | — |
Two caveats apply to every row, as they do to every Claude tier: "from" is the cheapest live route, not the only price, and route supply moves. The card is authoritative rather than this table.
What the prediction market is pricing
If you want a number that is neither a vendor statement nor an anonymous claim, the closest thing is a market where being wrong costs money.
The Polymarket event Next Claude Opus released by...? carried roughly $237,000 in total volume when we read it on September 21, about $117,000 of it traded in the previous week.
| Released by | Market-implied probability | Volume in that bucket |
|---|---|---|
| September 20 | 0.5% | $17,147 |
| September 24 | 79.5% | $22,418 |
| September 27 | 88.5% | $27,356 |
| September 30 | 90.0% | $96,257 |
| October 31 | 96.0% | $14,755 |
| December 31 | 98.1% | $5,283 |
Read those as prices, not prophecies. The crowd is putting roughly a one-in-five chance on nothing arriving by September 24 and roughly one-in-ten on nothing by September 30. That is conviction, not certainty — and the gap between those two numbers is the launch-week risk you are being asked to price.
The market's own resolution rules are the most useful sentence on the page. A qualifying model must be "launched and publicly accessible, including via open beta or open rolling waitlist signups" — and, explicitly, "a closed beta or any form of private access will not suffice." The same rules name Claude Opus 5.1, Claude Opus 5.5 and Opus 6 as qualifying. So the market is treating 5.5 as a live possibility while refusing to treat the testing reports as an event.
Internalise that before planning a migration: an early-access program is not a release. If Opus 5.5 appears in partner consoles and nowhere else, the market will still mark it unresolved — and so should your migration plan.
The capability claims
Beyond price, the September 20 posts claim three behavioural changes. None are in a changelog, and all three would be significant if they survive the launch:
- Thinking is "ALWAYS adaptive (no off mode)." If reasoning effort can no longer be disabled, cost control shifts from turning thinking off to choosing a cheaper tier — which is exactly the trade that makes the cache-read figure matter.
- "Forced tool use is retired." This would affect agent scaffolding that depends on tool-call guarantees rather than tool availability.
- Context reported at 872K, down from a reported 936K. A lower reported maximum on a version bump would be unusual, and it is the kind of number that is easy to get wrong from outside.
Treat these as day-one test items, not as a feature list.
What developers are posting
The screenshots circulating with the leak show 3D and front-end generations: a Waymo vehicle model, a Three.js car configurator, a page-sized interactive scene from a single prompt, and a claim of "one shot" JavaScript output.
These are impressions, not benchmarks. No prompt, no seed, no harness, no fixed baseline — and every one is a demonstration its author chose to publish. The same category of evidence declared Opus 5.2 imminent two weeks ago. Benchmark your own tasks when the model is real.
The same week, on the other side
Opus 5.5 is not arriving in a vacuum, and the collision is the story.
- GPT-6 Sol is the tier developers are waiting on from OpenAI — the "daily driver" under GPT-6 Astra, which stays the expensive frontier option. We track the routing evidence and the Claude-versus-GPT picture in GPT-6 Sol vs GPT-6 Astra and GPT-6 Sol API pricing.
- Gemini 4 Pro is the third front, with checkpoints reported in testing in the same window and no date from Google.
- Fable 5.2 appears in the same leak thread, with developers claiming early access — the same "partner access, not a release" pattern.
Three labs converging on one week is why a leak this thin travelled this far. It is also why every screenshot is starting to look like confirmation.
What to use while you wait
The decision does not depend on the leak being true.
- Strongest Claude available. Claude Fable 5.1, from $2.8944/M input against an official $10/M. Anthropic's own recommendation for demanding reasoning and long-horizon agentic work.
- Best value on agentic coding and enterprise workloads. Claude Opus 5 at $1.4847/M input — 70% below official, on 11 live routes. Anthropic's stated default for most workloads, and currently cheaper than the rumoured price of the model that would replace it.
- High-volume or latency-sensitive work. Sonnet 5 at $0.1875/M input, or Haiku 4.5 at $0.09/M. Sonnet 5's pricing became permanent on August 10, so it can be planned against rather than around.
- An exact older behaviour. Opus 4.7 and Opus 4.8, both at $0.45/M input from 13 live routes each — useful when your evaluation was built against their output distribution.
The one answer that is not on the list is "wait." A model with no announced date has no date, and every week of waiting is a week of not building the evaluation that would tell you whether the new tier is worth switching for.
When Opus 5.5 ships
Three things decide whether a launch costs you a migration or a config change.
- This URL keeps working. This page is the tracker for the Opus 5.5 generation. When Anthropic publishes a real model ID, price and limits, they land here and the "not published" row gets replaced.
- Your key does not change. APIMaster exposes one OpenAI-compatible endpoint, so adopting a new Claude tier is a
modelstring change rather than a new account, SDK or billing relationship. The same key that callsclaude-opus-5will call Opus 5.5 once a verified route exists. - We will not list it before it is verifiable. A model ID with no official price, no model card and no observable behaviour is a placeholder, and a placeholder returning tokens from an unknown source is worse than no listing. Once a route passes fingerprint checks and the official ID exists, it is added and priced on the Claude models page.
How to check an Opus 5.5 route when one appears
Launch week is when relabelling spikes: channels serving an older or smaller model and printing the new flagship's name. The label is not evidence.
- A response naming Opus 5.5 proves the client or gateway was configured to print that string. The
modelfield is filled in by the serving party, and a relay can set it independently of what generated the tokens. Our verification guide works through that distinction from the buyer's side. - A catalogue slug proves someone prepared an entry, not that the entry serves traffic.
- A leaked price proves someone typed a price.
- A partner-program screenshot proves access existed, not that a public model does.
What discriminates is behaviour: known-answer questions the new tier should handle differently from the tier below it, a fixed task battery run against both, and a fingerprint check before you move production traffic. Our model tester covers the last step, and it is free.
Run it before you switch, not after. The discount you get from a route that is not serving the model you ordered is not a discount.
FAQ
Has Anthropic released Claude Opus 5.5?
No. As of September 21, 2026 there is no Anthropic announcement, model ID, published price or release date, and the name appears nowhere in Anthropic's public documentation.
Is the next Claude Opus 5.2 or 5.5?
Publicly, nobody outside Anthropic knows. Expectation moved decisively to 5.5 on September 20, based on a leaker who attributes the testing build to the codename claude-wafer-eap. Earlier 5.2 routing reports are consistent with a test checkpoint; they are not evidence that 5.2 was cancelled.
What is claude-wafer-eap?
The codename the September 20 leak attributes to the testing build, with "EAP" presumably standing for early access program. Anthropic has not confirmed it. An entry in an early-access program is not a public release.
Is the $4 per million input price for Opus 5.5 real?
Not verified, and the source is weaker than the figure's reach suggests: the number comes from a third-party graphic stamped Unofficial concept · Unconfirmed details. The direction — a cheaper input price and much cheaper cache reads — is plausible and consistent with how Fable 5.1 shipped. The figure is not a price until it appears on a vendor page.
When will Opus 5.5 be released?
Nobody has announced a date. The leak said Tuesday, September 22, while allowing it could move earlier to Monday, September 21. Prediction markets put roughly 80% on a release by September 24 and roughly 90% by September 30.
Should I wait for Opus 5.5 before building?
No. An unannounced model has no date, no price and no verifiable route. Claude Opus 5 and Claude Fable 5.1 are available now with stable IDs, and migrating to a future tier is a model-string change on the same key.
Will APIMaster list Opus 5.5?
As soon as there is a published model ID and at least one channel that passes fingerprint checks — the same bar every other Claude tier on the site cleared. Until then there is nothing to sell, and inventing a listing would be the opposite of useful.
Is Opus 5.5 cheaper than Opus 5?
If the leaked numbers hold, its sticker price is 20% lower at $4 versus $5. Today, the cheapest live Opus 5 route on APIMaster is $1.4847 per million input tokens — about 63% below the rumoured 5.5 price. "Cheaper than Opus 5" is true of the sticker and false of what you can actually pay right now.
Sources and verification date
- Primary posts, read directly on X on September 21, 2026, with the timestamps quoted above: @lyraxana, September 20, 13:43:50 UTC, @synaptized, 15:30:41 UTC, @lyraxana, 15:35:40 UTC, @chetaslua, 16:15:56 UTC. Same account, earlier calls: Gemini 4 Pro checkpoint, September 17, Nano Banana 2.5, September 18.
- Polymarket — Next Claude Opus released by...?: volumes and implied probabilities read through the public Polymarket API on September 21, 2026, including the market's own rules on what counts as a release.
- Anthropic — Models overview and Claude API release notes, checked September 21, 2026: current lineup, model IDs, official pricing, positioning. Neither page references Opus 5.5,
claude-opus-5-5,claude-wafer-eapor Fable 5.2. - APIMaster live marketplace routes and the public Claude pricing table, checked September 21, 2026. Route availability and prices change after publication.
Start with Claude Opus 5 on APIMaster.ai
As of September 21, 2026, Opus 5.5 is a codename in a screenshot. Claude Opus 5 is a model you can call in the next five minutes, at $1.4847 per million input tokens — roughly 70% below Anthropic's official price — with Fable 5.1 on the same key when a task needs the frontier tier.
Create an APIMaster account, generate a key in the console, point your client at https://apimaster.ai/v1 and set the model to claude-opus-5. Every Claude channel is fingerprint-checked before it is listed, and the same endpoint covers Claude, GPT, Gemini, DeepSeek and Kimi — so when Opus 5.5 becomes a real model ID, the upgrade is one string in your config.