Direct answer
Claude Opus 5 is available now, but changing production traffic is optional
Anthropic launched claude-opus-5 on July 24 with a 1-million-token context window, 128,000-token maximum synchronous output, and the same standard list price as Claude Opus 4.8: $5 per million input tokens and $25 per million output tokens. Anthropic lists the model on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry. Do not treat availability as a forced migration. Opus 4.8 remains active, and the safer default is to test Opus 5 against an owned workload before changing production traffic.
The immediate compatibility checks are concrete: use the correct surface-specific model ID, remove non-default sampling parameters, and do not disable thinking at effort xhigh or max. Those request combinations return HTTP 400 errors under Anthropic’s current documentation.
Decision first
Stay, evaluate, or migrate
Opus 4.8 already works
Keep the active pinned model when the workload is stable, the launch adds no measured business value, or a partner surface has not exposed Opus 5 in your region or tenant.
You own a difficult workload
Run a fixed eval when complex agentic coding, long-context work, tool use, or enterprise analysis is important enough to justify a controlled comparison.
The measured result wins
Change traffic only after quality, latency, token use, failure rate, and total cost meet written thresholds and the old model remains available as a rollback path.
What changed July 24
One new pinned model ID, wider effort control, and two easy-to-miss request failures
claude-opus-5. Anthropic describes dateless model IDs in the current generation as pinned snapshots, not evergreen aliases that silently advance.Anthropic positions Opus 5 for complex agentic coding and enterprise work and calls it a “step-change” over Opus 4.8. That is a provider claim, not an independent TPTK benchmark. The decision-relevant facts are the published interface, price, lifecycle, and failure conditions.
The model supports effort levels low, medium, high, xhigh, and max, with high as the documented default on the Claude API and Claude Code. Thinking is on by default. If an existing request tries to combine disabled thinking with xhigh or max, Anthropic says the API returns a 400 error. Non-default temperature, top_p, or top_k values also return a 400 error on Claude 4.7 and later.
Two launch-adjacent features remain beta: mid-conversation tool changes require the mid-conversation-tool-changes-2026-07-01 header, while default server-side fallback behavior requires the server-side-fallback-2026-07-01 header. Treat both as separate adoption decisions rather than evidence that every Opus 5 request needs a beta header.
Reproducible cost check
A fixed 10M-input, 2M-output workload costs $100 at Opus 5 standard list price
The scenarios below multiply Anthropic’s published per-million-token prices by the same hypothetical workload: 10 million uncached input tokens and 2 million output tokens. They are planning arithmetic, not quotes.
Boundary: The examples exclude prompt caching, Message Batches discounts, tool fees, taxes, contracts, partner-cloud pricing differences, and workload-specific tokenization. Fast mode is a first-party Claude API research preview and is not available on partner-operated cloud platforms. US-only inference uses a 1.1x multiplier; global routing is the standard-price default.
The same per-token price does not guarantee the same invoice. Output length, reasoning-token use, retries, cache hits, tool calls, and success rate can change total spend. Measure tokens and completed-task cost on your own eval rather than assuming price parity means workload parity.
Availability map
Separate available now from gradual rollout and research preview
Available now
Use claude-opus-5. Anthropic documents 1M context, 128k synchronous maximum output, adaptive thinking, and standard $5/$25 token pricing.
Available now
Use anthropic.claude-opus-5 on the Messages-API Bedrock surface. Confirm regional access and the endpoint type in your account before rollout.
Available now
Anthropic lists the Google Cloud ID as claude-opus-5. Endpoint geography and project access remain tenant-specific.
Available now
Anthropic lists Microsoft Foundry as a launch surface. Verify catalog access, region, quota, and account pricing before production use.
Gradual rollout
GitHub says Opus 5 is rolling out to Pro+, Max, Business, and Enterprise across its supported clients. Organization administrators may need to enable the model policy.
Research preview
Fast mode is first-party Claude API only at $10/$50 per million input/output tokens. It is unavailable on Claude Platform on AWS and partner-operated clouds.
“Available” is not the same as visible in every tenant. Partner catalogs can vary by region, plan, quota, organization policy, and rollout timing. The practical check is to confirm the exact model ID in the surface where the workload will actually run.
Five-check migration plan
Test compatibility, cost, and rollback before changing traffic
- 1
Freeze a representative eval set
Use real task shapes with sensitive data removed: long-context analysis, tool calls, coding changes, structured outputs, refusal paths, and the cases that currently fail or require human correction. Record the old model and settings.
- 2
Map the surface-specific ID
Use
claude-opus-5on the Claude API and Google Cloud, oranthropic.claude-opus-5on Bedrock. Do not assume a cross-cloud alias or a legacy integration will resolve the same identifier. - 3
Remove incompatible request settings
Omit non-default sampling parameters. If thinking must be disabled, keep effort at
highor lower. Add beta headers only for the beta feature being tested. - 4
Measure completed-task economics
Compare quality, latency, input tokens, output and thinking tokens, tool calls, retries, human-review time, and successful completions. Standard token rates match Opus 4.8, but workload cost can still differ.
- 5
Set rollout and rollback thresholds
Start with bounded traffic, preserve Opus 4.8 as the fallback while it remains active, monitor 400s and refusals, and write the condition that returns traffic to the prior pinned model.
A new model does not retire Opus 4.8
Anthropic lists claude-opus-5 as active, with tentative retirement not sooner than July 24, 2027.
Anthropic also lists claude-opus-4-8 as active, with tentative retirement not sooner than May 28, 2027. No forced upgrade is documented.
Fast mode was removed July 24. Requests to Opus 4.7 with speed: "fast" now error, while Opus 4.7 remains available at standard speed.
That distinction changes the action. Teams on Opus 4.8 can evaluate without urgency. Teams still sending fast-mode requests to Opus 4.7 have a live compatibility problem and should remove the speed setting or move the workload to a supported fast-mode model after testing.
What is confirmed and what still requires your own test
Model IDs, context and output limits, standard and fast pricing, documented cloud surfaces, lifecycle states, and listed 400-error conditions come from current official documentation.
Anthropic’s “step-change” language and capability positioning describe its own model. TPTK has not independently benchmarked Opus 5 and does not present provider benchmarks as guaranteed results.
Quality, latency, token use, tool reliability, regional availability, quota, and completed-task cost must be measured in the reader’s own surface and workload.
Implementation path
Need an eval or migration plan, not another model announcement?
TPTK can help a small business or professional team freeze a representative test set, compare outputs and completed-task cost, define human review, and document rollout and rollback rules.
Claude Opus 5 FAQ
How much does Claude Opus 5 cost?
Anthropic lists standard Claude Opus 5 API pricing at $5 per million input tokens and $25 per million output tokens, the same rates as Claude Opus 4.8. Fast mode costs $10 and $50 respectively, and US-only inference applies a 1.1x multiplier.
What is the Claude Opus 5 model ID?
The Claude API and Google Cloud ID is claude-opus-5. Amazon Bedrock uses anthropic.claude-opus-5. Anthropic describes current dateless IDs as pinned snapshots rather than evergreen aliases.
Do I have to migrate from Claude Opus 4.8?
No. Anthropic lists both Claude Opus 5 and Claude Opus 4.8 as active, and it has not documented a forced Opus 4.8 migration. Evaluate Opus 5 on an owned workload before changing production traffic.
Which Claude Opus 5 request settings can return a 400 error?
Anthropic says disabling thinking at effort xhigh or max returns a 400 error on Opus 5. Non-default temperature, top_p, or top_k values also return a 400 error on Claude 4.7 and later.
Update log
July 28, 2026 at 5:16 AM PDT: Initial publication after a missed-interval continuity audit recovered Anthropic’s July 24 launch. Model IDs, specifications, prices, partner surfaces, lifecycle states, request failures, GitHub rollout, cost arithmetic, images, category, links, and schema were verified.
Sources
- Anthropic: Claude Platform release notes. July 24, 2026 entry; accessed July 28, 2026.
- Anthropic: Models overview. Accessed July 28, 2026.
- Anthropic: Claude pricing. Accessed July 28, 2026.
- Anthropic: Model deprecations. Accessed July 28, 2026.
- Anthropic: What’s new in Claude Opus 5. Accessed July 28, 2026.
- GitHub: Claude Opus 5 in GitHub Copilot. July 24, 2026; accessed July 28, 2026.
All central factual claims use first-party sources. TPTK’s cost examples multiply published list prices by a disclosed hypothetical workload. No firsthand benchmark claim is made.
Editorial and monetization note: No Anthropic affiliate route or audience-matched verified Amazon subscription bounty fits this enterprise API decision, so none appears. The implementation link is a TPTK service path. Material corrections and source changes will be recorded in the update log.