Claude Opus 5
Opus 5 went into production on 24 July 2026 and holds rank one on the WebDev Arena leaderboard today, at a rate of $5 in and $25 out per million tokens. For coding workloads and multi-step agentic pipelines, it is the first pick in this reference; no official deployment route exists from Iran.
- rank one on WebDev Arena
- 1M token context window
- knowledge cutoff May 2026
Last checked: This is a reference page. It is re-checked against the vendor sources and updated when a new version ships. Change log
What changed in the architecture relative to 4.8
Two components with a measurable effect on output are the knowledge cutoff and agentic behaviour. Opus 5 is provisioned with a knowledge cutoff of May 2026, where Opus 4.8 was fixed at January 2026. On workloads tied to recent libraries or a framework that shipped changes lately, that gap shows up in real output.
The pricing architecture is unchanged: the same $5 and $25 as Opus 4.8. That means the migration path from 4.8 to 5 carries no added unit cost, a condition that recurs less often than it should in this release cycle.
Why the announcement figures are not in our table
Anthropic published no absolute benchmark values for Opus 5. The announcement references Frontier-Bench, ARC-AGI 3, Zapier AutomationBench and OSWorld, but every reported value is relative: "more than double Opus 4.8", "three times the nearest competing model", "within half a percentage point of Fable 5's peak at half the cost per task". None of that can be entered into a column that requires an absolute figure, so the Frontier-Bench field in our coding-category table stays marked unverified.
That is a constraint of our measurement method, not a deficiency in the model. If Anthropic publishes an absolute figure, it is added to the table the same day and may well shift the ranking.
The fast-mode configuration, a component that gets documented less than it should
Opus 5 ships with a fast-mode configuration that, at twice the base price, increases response throughput by roughly 2.5x. For interactive workloads where a user is waiting on output, that trade-off usually pays off. For an overnight batch pipeline, it never does. Any team costing a service against consumption needs to model the two configurations separately, or the invoice doubles with no traceable cause.
Specifications
Every number here comes from the vendor's own page, with that page linked beside it.
- API model ID
- claude-opus-5 Source
- Released
- 24 July 2026 Source
- Context window
- 1 million tokens Source
- Max output
- 128k tokens Source
- Input price
- $5 per million tokens Source
- Output price
- $25 per million tokens Source
- Knowledge cutoff
- May 2026 Source
- Open weights
- no Source
- Input
- text, image Source
- Output
- text Source
- Where it ships
- Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry Source
Using it from Iran
This column carries a date and says how it was checked, because most listicles guess it. Where we have only read a vendor policy page, the note below says exactly that.
- Reachable
- blocked
- Payment
- no working route
- Free tier
- no
- Measured on
How we checked: Iran is registered on neither of Anthropic two supported-countries lists, at the API layer or the claude.ai layer. That status is extracted from Anthropic own policy document, not from a direct network measurement; an independent network test has not been run yet, and its result will be logged here once it is.
Documented strengths
- Rank one on WebDev Arena at a score of 1692, ahead of Kimi K3 Max and Qwen3.8 Max
- The most recent knowledge cutoff across the entire Anthropic model fleet: May 2026
- Identical pricing architecture to Opus 4.8, so the migration path carries no added unit cost
- Full coverage across all four deployment routes: Anthropic's own API, Bedrock, Google Cloud and Microsoft Foundry
Known weaknesses and limits
- Anthropic published no absolute benchmark values, leaving part of the table fields empty for this model.
- No official deployment route from Iran: it is absent from the supported-countries list and Iranian card payment processing is not operational.
- The fast-mode configuration costs twice the base price and, without monitoring, will silently double an invoice.
- On the general text leaderboard it ranks seventh, below several older Anthropic models, so it cannot be classified as optimal across every workload.
Technical verdict
For coding workloads, this is the first pick in this reference today, and the table data confirms it. But if your workflow is general conversation rather than code, the price premium is not justified: on the text leaderboard, its gap to cheaper options sits within statistical noise.
Frequently raised questions, with documented answers
From an architecture standpoint, is Opus 5 superior to Fable 5
For coding workloads, yes. On WebDev Arena, Opus 5 scores 1692 against Fable 5's 1628, at half the output rate. Fable 5 holds rank one on the general text leaderboard, so for conversation and long-form non-coding work it still sits higher.
Does migrating from Opus 4.8 have technical justification
If your workload is code, or requires knowledge past January 2026, yes, because the unit price is unchanged. If you have a stable system built on 4.8 that does not depend on recent knowledge, there is no urgency to migrate.
What is the access path from Iran to this model
No official route exists. Iran is not registered on Anthropic's supported-countries list and Iranian card payment processing is not operational. This reference does not propose circumvention methods and does not sell accounts.