Anthropic
Claude Opus 4.8 (Adaptive Reasoning, Max Effort)
Released 
May 2026

Anthropic
Claude Opus 4.8 (Adaptive Reasoning, Max Effort)

Anthropic's flagship for long-horizon agents that flag their own mistakes and skip fewer required tool calls.

Modality:
Text
Image
PDF
model ID
anthropic/claude-opus-4.8

Output Speed *

0.00
tok/s

Intelligence Index *

42
/ 100

Context Window *

1000000
tokens

Input price

30
Anytoken

Output price

150
Anytoken
Claude Opus 4.8: A Flagship Built for Autonomous Agents That Own Their Own Errors Claude Opus 4.8 is Anthropic's flagship model, sitting above Sonnet and Haiku in the Claude family. Released May 28, 2026, it builds on Opus 4.7 with gains across coding, agentic tasks, and knowledge work. Its defining change is judgment rather than raw capability: Anthropic reports Opus 4.8 is roughly four times less likely than Opus 4.7 to let flaws in its own code pass unremarked, and it triggers required tool calls more reliably. It targets long-running, unattended agentic engineering and multi-step coding where sustained coherence matters most. Start building with Claude Opus 4.8 through the AnyAPI.ai API.

Performance

Where Opus 4.8 Separates Itself: Honesty in Long-Horizon Agentic Work

Claude Opus 4.8 is strongest on sustained, unattended agentic engineering: long coding runs, tool orchestration, and multi-step debugging where the model must stay coherent across many turns. Anthropic reports it is roughly four times less likely than Opus 4.7 to let flaws in its own code pass unremarked, and it scores 69.2% on SWE-bench Pro versus 64.3% for its predecessor. This matters because agentic pipelines fail silently when a model declares victory early. In production, a model that proactively flags uncertainty and triggers required tool calls reduces the human review burden on autonomous, long-running workflows.

Benchmarks

Opus 4.8 on Coding, Terminal, and Knowledge-Work Benchmarks

On SWE-bench Pro—drawn from actively maintained repositories with no public ground-truth leakage—Opus 4.8 reaches 69.2%, roughly five points ahead of Opus 4.7 and over ten points ahead of GPT-5.5 (58.6%) and Gemini 3.1 Pro (54.2%) as reported in the model's system card. On Terminal-Bench 2.1, covering real command-line tasks, it scores 74.6% versus 66.1% for Opus 4.7. Independent analysis notes the pattern that its lead widens on harder, less contaminated benchmark variants, and that at minimum effort Opus 4.8 already matches Opus 4.7's peak SWE-bench Pro performance at maximum effort.

Output Speed

*
0.00
tok/s

Intelligence Index

*
42
/ 100

MMLU *

Broad world knowledge and problem-solving
0
%

GPQA *

PhD-level scientific reasoning across physics, biology, chemistry.
92
%

HLE *

Adherence to multi-step structured instructions.
49
%

LiveCodeBench *

Tool-calling reliability in long agentic loops.
0
%

Technical Specifications

What the model supports

Claude Opus 4.8 accepts text, image, and file input and returns text. It ships with a 1M-token context window by default and up to 128K output tokens (300K on the Batches API via beta header). The highest-impact specifications for production teams are its reasoning model: adaptive thinking lets the model decide per turn whether to reason, and the effort parameter (default: high) trades thoroughness against token efficiency. Because thinking tokens bill as output, and the default moved from Opus 4.7's prior behavior, teams migrating should audit effort settings before scaling to control per-call cost.
Verified Specifications — 
Claude Opus 4.8 (Adaptive Reasoning, Max Effort)
*
Input modalities
Text
Image
PDF
output modalities
Text
Context window
1000000
 tokens
Maximum output tokens
128000
Reasoning
Yes
Knowledge cutoff
May 2026
Pricing (standard)
30
 AnyTokens in
 / 
150
 AnyTokens out

Comparison

Claude Opus 4.8 vs Claude Opus 4.7: What Actually Changes?

Opus 4.8 and Opus 4.7 are natural comparison points: they share the same 1M-token context window, 128K max output, January 2026 knowledge cutoff, and tool surface, and Anthropic priced 4.8 identically to its predecessor. The decision is not about new capabilities—Anthropic itself called 4.8 "a modest but tangible improvement." The real change is reliability. Opus 4.8 introduces adaptive thinking, fixes tool-calling and comment-verbosity issues reported on 4.7, and is roughly four times less likely to let flaws in its own code pass unremarked.

Dimension
Claude Opus 4.8 (Adaptive Reasoning, Max Effort)
Claude Opus 4.7 (Adaptive Reasoning, Max Effort)
Context window *
1000000
tokens
1000000
tokens
Output speed *
0.00
tok/s
0.00
tok/s
Intelligence Index *
42
40.7
Input pricing
30
AnyToken
30
AnyToken
Output pricing
150
AnyToken
150
AnyToken
Knowledge cutoff *
May 2026
April 2026

Choose Opus 4.8 when you run long, unattended agentic sessions, code migrations, or multi-step debugging where silent failures and skipped tool calls are expensive, and when honest self-flagging reduces human review. It is also the drop-in upgrade at the same token price. Choose Opus 4.7 only if you have already validated and pinned it in production and cannot spare the time to re-test adaptive thinking and the new default effort behavior, since migrating may shift per-call token consumption unless you set effort explicitly.

Limitations & Trade-offs

Where Claude Opus 4.8 (Adaptive Reasoning, Max Effort) falls short

1
Incremental raw capability. Anthropic itself described Opus 4.8 as "a modest but tangible improvement" over Opus 4.7, and independent reviewers agree the gains are reliability-driven rather than a new capability ceiling. If your workload already runs well on Opus 4.7 and doesn't suffer from silent errors or skipped tool calls, the upgrade may not change output quality much. Teams expecting a step-change should look to newer Opus-tier releases instead.
2
Effort and thinking cost drift. The effort parameter defaults to high, and adaptive thinking spends thinking tokens—billed as output—whenever a turn needs them. This means per-call cost in your dashboard can move even though headline pricing is unchanged from Opus 4.7. Workloads migrating from 4.7 should run a representative sample and set effort: "low" explicitly for high-volume, straightforward calls before scaling, or costs on bimodal pipelines may creep upward.
3
Premium tier for high-volume routine work. Opus 4.8 sits at the top of the Claude family's cost tier. For classification, RAG responses, routine content generation, and simple tool use, a Sonnet-class model is materially cheaper per token and generally the better default. Reserve Opus 4.8 for high-stakes engineering and long-horizon agentic work where its judgment and coding advantage justify the higher cost.
4
January 2026 knowledge cutoff and preview Fast mode. Opus 4.8 carries the same January 2026 knowledge cutoff as Opus 4.7, so questions about very recent libraries or APIs may need retrieval or tool augmentation. Additionally, Fast mode—which runs at roughly 2.5x speed—remains in research preview; availability can vary by access tier, so it should not be a hard production dependency for latency-sensitive systems until it stabilizes.

Best-Fit Workloads

Where this model earns its place

01

Autonomous long-horizon coding agents


Opus 4.8's core strength is sustaining agentic engineering unattended. It triggers required tool calls more reliably than Opus 4.7, handles compaction better so long traces stay on task, and is roughly four times less likely to let its own code flaws pass unremarked. For products running multi-hour agent sessions—codebase migrations, multi-stage debugging, asynchronous pipelines—this reduces silent failures and the human oversight burden. The 1M-token context lets large codebases fit in a single session.

02

Large-scale codebase migrations and refactors


With a 1M-token context window and 128K max output, Opus 4.8 can hold substantial portions of a codebase in context and emit long artifacts without truncation. Anthropic's dynamic workflows feature (in Claude Code) fans a hard problem across parallel subagents to handle migrations spanning hundreds of thousands of lines. Its 88.6% SWE-bench Verified and 69.2% SWE-bench Pro scores support multi-file changes that respect architecture rather than making local, brittle edits.

03

Professional knowledge work and analysis


Opus 4.8 maintains quality across very long outputs for drafting documents, building presentations, and analyzing data. Enterprise testers reported higher-quality, more information-dense analysis than prior Opus models, with a notable tendency to proactively flag issues in inputs and outputs that other models leave for users to catch. On GDPval-AA, which measures economically valuable knowledge work, Anthropic reports Opus 4.8 leading the frontier cluster—useful for high-value professional deliverables where errors are costly.

04

Agentic computer use


Opus 4.8 scores 83.4% on OSWorld-Verified, ahead of GPT-5.5 (78.7%) and Gemini 3.1 Pro (76.2%) as reported, making it a practical fit for agents that operate a computer interface across many steps. This capability underpins running parallel subagents that verify their own work before reporting back. As with all agentic deployments, treat honesty and self-flagging as a safety layer rather than a substitute for verification on high-stakes actions.

Pricing in anytokens via AnyAPI
Input
30
Output
150
Cache write
37.5
Cache read
3

Integration

Access Claude Opus 4.8 (Adaptive Reasoning, Max Effort) via AnyAPI.ai

Access Claude Opus 4.8 (Adaptive Reasoning, Max Effort) through AnyAPI.ai using a unified API built for multi-model AI applications. Integrate Claude Opus 4.8 (Adaptive Reasoning, Max Effort) without maintaining a separate provider-specific connection, and keep the flexibility to test, switch, or combine models as your application requirements evolve.

01

One API integration

Access Claude Opus 4.8 (Adaptive Reasoning, Max Effort) and other AI models through the same API workflow instead of maintaining separate integrations for every provider.

02

Easy model switching

Test Claude Opus 4.8 (Adaptive Reasoning, Max Effort) against alternative models or switch models as your performance, capability, or cost requirements change without rebuilding your application around another provider API.

03

Flexible for production

Use Claude Opus 4.8 (Adaptive Reasoning, Max Effort) from experimentation through production while keeping your AI stack flexible as workloads, traffic, and model requirements evolve.

04

Multi-model applications

Use Claude Opus 4.8 (Adaptive Reasoning, Max Effort) for the workloads where it performs best and combine it with other models for tasks that require different capabilities, performance, or efficiency.

Frequently Asked Questions

Answers to common questions about integrating and using this AI model via AnyAPI.ai

Claude Opus 4.8 has a 1,000,000-token context window, available by default on the Claude API, Amazon Bedrock, Google Vertex AI, and Microsoft Foundry. It supports up to 128,000 output tokens on the synchronous Messages API, and up to 300,000 output tokens on the Message Batches API using a beta header. Both figures are unchanged from Opus 4.7.

Opus 4.8 improves on Opus 4.7 across coding benchmarks: 69.2% on SWE-bench Pro (up from 64.3%) and 74.6% on Terminal-Bench 2.1 (up from 66.1%), with 88.6% on SWE-bench Verified. The larger change is reliability—Anthropic reports it is about four times less likely to let flaws in its own code pass unremarked, and it fixes tool-calling issues seen in 4.7.

Yes. Opus 4.8 supports adaptive thinking, where the model decides per turn whether to reason, enabled with thinking: {type: "adaptive"}. Thinking is off unless you set it explicitly. A separate effort parameter (default: high) controls how thoroughly the model works. Thinking tokens bill as output tokens, so audit effort settings to manage cost on high-volume workloads.

Claude Opus 4.8 has a knowledge cutoff of January 2026, the same as Opus 4.7. For questions about very recent libraries, APIs, or events after that date, pair the model with retrieval or web/tool augmentation rather than relying on its training data alone.

Claude Opus 4.8 accepts text, image, and file/document input, and produces text output. It supports function calling via tools and tool_choice, with improved tool triggering over Opus 4.7. Its API model identifier is claude-opus-4-8. It does not generate images, audio, or video.

* Benchmark data source: Artificial Analysis artificialanalysis.ai