Claude Opus 4.8 is Anthropic's May 2026 model for demanding coding, browser agents, tool use and long-context knowledge work. It is still a capable choice for a controlled pilot, but there is an important qualification: the current Claude Platform catalog labels it Legacy and recommends considering Opus 5.
That lifecycle fact is the decision anchor as of September 20, 2026. The catalog still lists the same regular price as Opus 4.7—$5 per million input tokens and $25 per million output tokens—along with a 1M context window, 128K maximum output and adaptive thinking. A new integration should therefore ask two questions at once: does 4.8 solve this task, and is its support horizon worth adopting? Start with the Claude Opus 4.8 model resource, then confirm the live route.
Key takeaways
Opus 4.8 is the 4.7 successor, with effort controls, more deliberate tool use, Fast Mode and Dynamic Workflows in Claude Code.
The official catalog lists
claude-opus-4-8, 1M context, 128K output, adaptive thinking, high default effort and a January 2026 knowledge cutoff.Independent early evaluations put it at the top of one composite index, but Terminal-Bench 2.1 still favoured GPT-5.5 in the cited harness. Benchmark labels matter.
Higher effort, thinking, long context and retries can change completed-task cost. A $5/$25 token snapshot is not a task budget.
It remains available, but a new production buyer should compare Opus 5 and document a migration plan.
Claude Opus 4.8 at a glance
| Question | Catalog snapshot checked 2026-09-20 | Decision boundary |
|---|---|---|
| Stable API ID | claude-opus-4-8 | Pin the exact ID in logs; provider aliases can differ. |
| Release and status | May 28, 2026; Active (legacy) | Available does not mean current family endpoint. |
| Context / maximum output | 1M tokens / 128K tokens | Catalog ceilings; client and provider windows may be smaller. |
| Input / output | Text and images / text | Tools and browser permissions come from the route. |
| Thinking and effort | Adaptive; default high; extra/xhigh and max documented | Compare effort with the task, not only model names. |
| Standard API price | $5 input / $25 output per MTok | Cache, Batch, Fast Mode, provider and plan terms are separate. |
| Knowledge cutoff | January 2026 | Current facts require retrieval or source checks. |
| Retirement | Not sooner than May 28, 2027 | Ask whether a new build should target Opus 5 instead. |
The prompt resources and review resources are the right place for source-level entries. This overview keeps the lifecycle and workload decision in view.
What changed from Opus 4.7?
Anthropic says 4.8 improves judgment, honesty about unfinished work, instruction-following and tool efficiency. Its launch page reports that the model was around four times less likely than its predecessor to let flaws in generated code pass without comment. That is a vendor evaluation claim, not a promise that a repository will be safe without tests.
Three product changes matter in practice:
Effort control: Anthropic says high is the default balance, while extra or
xhighand max spend more tokens on difficult or asynchronous work. Lower effort can respond faster and consume rate limits more slowly.Fast Mode: Anthropic describes the same model running at roughly 2.5× speed, priced at $10 input and $50 output per million tokens. That is a route and price choice, not a new model ID.
Dynamic Workflows: In research preview for Claude Code Enterprise, Team and Max plans, Claude can plan a large job, run hundreds of parallel subagents and verify outputs before reporting back.
| Change | What Anthropic says | What you should verify |
|---|---|---|
| Agent follow-through | More independent work, clearer self-checking and better judgment | Whether your task has a stop condition, tests and permissions. |
| Tool calling | Fewer steps for the same intelligence in the launch evidence | The actual provider tools, rate limits and logs. |
| Effort | High default, plus extra/xhigh and max; lower effort is faster | Tokens, retries and completed-task cost at each setting. |
| Large workflows | Dynamic Workflows can coordinate hundreds of subagents | Plan eligibility, research-preview status and rollback controls. |
| Safety | Lower misaligned-behaviour rates than 4.7 in Anthropic's assessment | Application safeguards and human approval still remain necessary. |
The version change is not a simple “4.8 always wins” story. A long-term Reddit user wrote that 4.8 was a clear update for Playwright, Cloud CLI and Kubernetes CLI work, but felt GPT-5.5 was more autonomous on some very large codebases (original report). Another early thread contains a blunt counterexample: a user said 4.8 gave quick, confident answers with issues during broken-code repair (discussion). These reports establish task-specific experience, not a general ranking.
Benchmarks: read the harness before the number
The independent early-evaluation roundup gives a useful split. Artificial Analysis's composite Intelligence Index placed Opus 4.8 at 61.4 versus GPT-5.5 at 60.2, while Terminal-Bench 2.1 showed Opus 4.8 at 74.6 versus GPT-5.5 at 78.2. The same roundup reported SWE-Bench Pro at 69.2 for 4.8 versus 64.3 for Opus 4.7, and Finance Agent v2 at 53.9 versus Gemini 3.5 Flash at 57.9.
These numbers use different tasks, dates and harness choices. The independent review of Opus 4.8 also notes that Anthropic's SWE-bench figures were not independently reproduced, while its Artificial Analysis snapshot measured 84.6% on its own Terminal-Bench 2.1 run on August 3, 2026. Do not merge 74.6 and 84.6 into one score; they are different runs.
An academic red-team study adds a different boundary. Across 7,826 harmful intents and four automated jailbreak families, its strongest adaptive tree-of-attacks broke Opus 4.8 on 11.5% of intents. That is a controlled adversarial result, not a normal-use failure rate, but it is enough to reject “the model's safety score makes application controls unnecessary.” The study does not justify reproducing harmful prompts here.
Lifecycle, access and pricing
The current Platform page lists Opus 4.8 as available but legacy, with a migration link to Opus 5 and retirement not sooner than May 28, 2027. Anthropic lists Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS as model routes. Claude.ai and Claude Code plans expose their own selectors, quotas and feature gates; do not infer them from the API catalog.
| Route or feature | Current published boundary | What it does not establish |
|---|---|---|
| Claude API | claude-opus-4-8, $5/$25 per MTok | That a plan or browser exposes the same limits. |
| AWS / Google Cloud / Microsoft Foundry | Provider-specific model IDs and terms | That quota, region or tools match Anthropic's route. |
| Claude.ai / Claude Code | Effort control and product-specific access | That subscription access is API credit. |
| Fast Mode | Research-preview route, about 2.5× speed, $10/$50 per MTok | That every plan or provider offers it. |
| Prompt caching | Cache writes and reads have separate prices; cache can reduce repeated-context cost | That a long task is automatically cheap. |
| Batch API | 50% discount on input and output in the catalog | That interactive tool use follows Batch semantics. |
| Dynamic Workflows | Research preview for Enterprise, Team and Max in Claude Code | That every user can run hundreds of subagents. |
The price is a usage boundary, not a prediction. A long high-effort run can add thinking tokens, cache reads, retries, tool calls and subagent work. Measure cost per accepted result. Do not mix API prices with Claude.ai or Claude Code subscription limits, provider markups, Fast Mode, or Tabbit availability.
For the adjacent family choices, compare the Claude Sonnet 5 overview and the Claude Fable 5.1 review separately; their effort, pricing and lifecycle evidence is not interchangeable with Opus 4.8.
Unknown risks and practical self-check
Before trusting an Opus 4.8 result, record the exact model ID, client, provider, region, date, effort and tools. Define the input files, expected output, acceptance test and stop condition. For code, run the existing tests and inspect the diff. For browser or terminal agents, log tool calls and require confirmation before irreversible actions. For research, require citations and a contradiction pass.
The community reports make the context risk concrete. One Reddit tracker claimed up to 900,000 cache tokens per turn with 4.8 thinking enabled, compared with 14,000–34,000 for 4.7; another large-codebase user described 100+ files and hundreds of thousands of tokens in one session (source). These are user measurements, not catalog limits. They are a reason to measure context and effort in your own harness, not a reason to rewrite the official 1M window.
What Tabbit can and cannot establish
This draft did not run a signed-in Claude Opus 4.8 task in Tabbit Browser. It therefore makes no claim about model-selector visibility, effective context, latency, provider routing, subscription access, tool permissions or cost in Tabbit. If 4.8 appears in your selector, start with a reversible task, record the visible model name and effort setting, and compare the output with a known acceptance result.
The agentic browser guide, browser automation guide and AI browser guide describe client-level boundaries. The Tabbit Browser overview is not a substitute for an API or provider test.
Tabbit BrowserVerdict
Claude Opus 4.8 remains a credible pilot for complex coding, tool-heavy analysis and browser-agent work when its effort and cost fit the task. Its 4.7 improvements are meaningful, but independent evidence is mixed by harness and the current catalog already calls it legacy. Use a fixed acceptance test, compare Opus 5 before a new production commitment, and keep human approval around irreversible actions.
Sources and further questions
Primary sources are Anthropic's launch announcement, Platform model page, pricing documentation and system card. Independent sources are Digital Applied, ThePlanetTools and the red-team study. The internal research ledger records the community URLs and collection limits.
Is Opus 4.8 still available?
Yes, the catalog lists it as active but legacy. It recommends considering Opus 5 and says retirement is not sooner than May 28, 2027.
Does 1M context mean every client exposes 1M?
No. It is a catalog model limit. A provider, plan, UI or tool wrapper can expose a smaller effective context.
Should I use high, extra or max effort?
Start with the lowest setting that meets your acceptance test. Higher effort can improve difficult work but can also increase tokens, waiting time and rate-limit use.
Is Fast Mode a different model?
Anthropic describes it as a faster route for the same Opus 4.8 model, with separate $10/$50 per-million-token pricing. Availability is provider and plan dependent.
Is Opus 4.8 safer than 4.7?
Anthropic reports lower misaligned-behaviour rates than 4.7, while independent red-team work still found an adaptive attack surface. Keep application controls and human approval.
Can Tabbit use Opus 4.8?
This draft did not verify a signed-in Tabbit session. Check your selector and treat one successful task as a local observation, not a platform guarantee.
FAQ
What is Claude Opus 4.8?
Claude Opus 4.8 is Anthropic's May 2026 Opus release for demanding coding, tool use, browser agents and long-context knowledge work. Its API ID is claude-opus-4-8 and its catalog page lists a 1M context window and 128K maximum output.
Is Claude Opus 4.8 still current?
It is still available, but the Claude Platform page checked on September 20, 2026 labels it Legacy and recommends considering Claude Opus 5. The page says retirement is not sooner than May 28, 2027.
How much does Claude Opus 4.8 cost?
Standard API pricing is $5 per million input tokens and $25 per million output tokens. Fast Mode is listed at $10 and $50; cache, Batch, provider and consumer-plan terms are separate.
What effort settings does Opus 4.8 support?
Anthropic documents adaptive thinking with high as the default effort, plus extra or xhigh and max for harder work. Lower effort can respond faster and consume rate limits more slowly, while higher effort can use more tokens.
How is Opus 4.8 different from Opus 4.7?
Anthropic describes better judgment, instruction-following, tool efficiency and self-checking, alongside effort control, Fast Mode and Dynamic Workflows. Independent early evaluations show a mixed picture by harness rather than a universal win.
Can I use Claude Opus 4.8 in Tabbit?
This draft did not run a signed-in Tabbit Opus 4.8 task, so it makes no claim about selector visibility, effective context, latency, cost or tools. Check the live selector and verify one reversible task in your account.