Brilliant. Honest.

#1 Intelligence Index. 3.7% sycophancy rate — 4× lower than any competitor. Claude Opus 4.8 on Yesmind.

Honesty — Claude Opus 4.8

Is this investment strategy guaranteed to double my money in 6 months?

No investment strategy can guarantee doubling your money in 6 months. Any claim that promises guaranteed returns of this magnitude is either misleading or fraudulent. I’d rather give you an honest assessment than tell you what you want to hear.

Sycophancy override — honest response

Sycophancy rate: 3.7% vs industry avg ~15%. Claude tells you what’s true, not what’s pleasant.

Extended Thinking — Adaptive

> Prove √2 is irrational

⟳ thinking (effort: high)

1 Assume √2 = a/b (coprime)

2 Then 2b² = a² → a is even

3 Let a = 2k → 2b² = 4k² → b² = 2k²

4 b is even → contradicts coprime

✓ Proof complete — contradiction established

USAMO 2026: 96.7% · Humanity's Last Exam: #1

61.4

Intelligence Index (#1)

88.6%

SWE-bench Verified

96.7%

USAMO 2026

3.7%

Sycophancy (lowest)

What is Claude Opus 4.8?

Claude Opus 4.8 is Anthropic’s most capable model to date — and the first to claim the #1 position on the Artificial Analysis Intelligence Index with a score of 61.4, surpassing GPT-5.5 Pro (60.2) and Gemini 3.1 Pro (57.0). Built for engineers, researchers, and enterprises that need a model they can trust not just to get answers right, but to flag when it doesn’t know. Opus 4.8 glosses over its own failures just 3.7% of the time — a 4× improvement over its predecessor.

The headline upgrades go far beyond honesty. Adaptive thinking lets the model auto-calibrate reasoning depth per task — from quick responses to extended chain-of-thought, configurable across five effort levels. Dynamic workflows let Claude plan a complex task, spawn hundreds of parallel subagents, and verify the output — enabling codebase-scale migrations that once required entire engineering teams. The 1 million token context window swallows entire codebases, legal contracts, and research corpora in a single pass.

For math and science, the leap is staggering: 96.7% on USAMO 2026 (up from 69.3% on Opus 4.7) and #1 on Humanity’s Last Exam. For agentic web tasks, 84% on Online-Mind2Web. And when speed and cost matter, Fast Mode delivers 2.5× the throughput at one-third the price — $10/$50 per million tokens versus the standard $15/$75. All at the same API surface, with no breaking changes.

61.4

Intelligence Index (#1)

88.6 %

SWE-bench Verified

84 %

Browser Agent (Mind2Web)

96.7 %

USAMO 2026 Math

Adaptive Thinking

Claude Opus 4.8 automatically scales reasoning depth to match task complexity. Five configurable effort levels — low, medium, high, xhigh, max — let you control the compute budget. Trivial requests get instant responses. Multi-step proofs and architectural decisions get full extended thinking. No tokens wasted on tasks that don’t need them.

Dynamic Workflows

Plan a task once, execute at scale. Claude Opus 4.8 can spawn hundreds of parallel subagents, delegate subtasks, verify results, and synthesize outputs into a coherent deliverable. Codebase-scale refactors, multi-source research synthesis, and batch document processing that used to take teams now run autonomously in Claude Code.

Computer Use & Browser Agent

84% on Online-Mind2Web — the highest score among frontier models for browser-based task automation. 100% completion rate on Anthropic’s internal Super-Agent evaluation. Claude Opus 4.8 can navigate the web, fill forms, interact with UIs, and complete multi-step computer tasks with no human in the loop.

Math & Scientific Reasoning

From 69.3% to 96.7% on USAMO 2026 — a 27-point leap that puts Claude Opus 4.8 in a class of its own for mathematical reasoning. #1 on Humanity’s Last Exam and leading on CritPt frontier physics. Scientific research, quantitative analysis, and theorem-level problem-solving now fall within reach of a single API call.

Honesty & Hallucination Resistance

35.9% hallucination rate on AA-Omniscience — the lowest among all frontier models, substantially below Google and OpenAI peers. Claude Opus 4.8 abstains when uncertain rather than confabulating confident answers. Sycophancy-under-pressure evals: 0%. If you need a model you can actually trust for high-stakes decisions, this is the standard.

Dynamic Workflows with parallel subagents

Computer use and browser agent automation

AI honesty and hallucination rate comparison

What Makes Claude Opus 4.8 Different

#1 Intelligence Index

61.4 on Artificial Analysis’s independent Intelligence Index — the most rigorous third-party ranking of frontier models. Claude Opus 4.8 beats GPT-5.5 Pro (60.2) and Gemini 3.1 Pro (57.0) in head-to-head evaluations, winning 67% of direct comparisons against GPT-5.5. Not a single benchmark cherry-picked by Anthropic — an independent composite score across dozens of tasks.

Lowest Hallucination Rate

35.9% hallucination rate on AA-Omniscience, substantially below every other frontier model. When Claude Opus 4.8 doesn’t know something, it says so. Zero score on sycophancy-under-pressure evals: the model won’t flip a correct answer because a user pushes back. For legal, medical, financial, and engineering work, this matters more than any benchmark score.

1M Token Context Window

One million tokens on Claude API, Amazon Bedrock, and Google Vertex AI (200K on Microsoft Foundry). That’s entire codebases, multi-volume research libraries, and year-long conversation histories in a single context. Maximum output: 128K tokens per call. Long-context retrieval quality remains competitive at the top of MRCR evaluations.

Fast Mode: 2.5× Speed, 3× Cheaper

Fast Mode costs $10 input / $50 output per million tokens — versus the standard $15/$75. Throughput increases 2.5×. For latency-sensitive agentic workloads, high-volume document processing, and cost-conscious production pipelines, Fast Mode makes Opus 4.8 competitive on price without switching to a smaller model.

Claude Opus 4.8 vs Frontier Models

Independent benchmark data across coding, reasoning, science, and pricing.

Features	Claude Opus 4.8 Anthropic	GPT-5.5 Pro OpenAI	Gemini 3.1 Pro Google	DeepSeek V4 Pro DeepSeek
Intelligence Index	61.4 (#1)	60.2	57.0	TBD
SWE-bench Verified	88.6%	~80%	~78%	80.6%
USAMO 2026 Math	96.7%	TBD	TBD	TBD
Hallucination Rate	35.9% (lowest)	~45%	~50%	TBD
Computer Use / Browser	84%	78.7%	TBD	TBD
Context Window	1M tokens	1M tokens	1M+ tokens	1M tokens
Input Price / M tokens	$5.00	$5.00	$2.00	$0.435
Output Price / M tokens	$25.00	$30.00	$12.00	$0.87
Fast / Turbo Mode	2.5× speed @ $10/$50	N/A	Thinking levels	N/A
Prompt Caching
Mid-conversation System Messages

Dynamic Workflows — Agentic AI at Scale

From single prompt to coordinated multi-agent execution

STEP 01

Plan

Claude Opus 4.8 receives a high-level objective — e.g. “Migrate this 200K-line codebase from Python 2 to Python 3.” It breaks the task into subtasks, identifies dependencies, and defines a verification strategy before touching a single file.

STEP 02

Spawn Subagents

Hundreds of parallel Claude instances spin up to handle independent subtasks simultaneously. File-by-file migration, test generation, import rewrites — each subagent works on its assigned scope without blocking the others.

STEP 03

Verify

Each subagent’s output is checked against success criteria — tests pass, types are correct, behavior is preserved. Failed verification loops back automatically. Claude’s 4× improved honesty means it surfaces real failures instead of marking them done.

STEP 04

Report

A synthesis agent aggregates results, resolves conflicts, and delivers a final coherent output — PR-ready code, summary report, or structured data. What took an engineering team days completes in a single orchestrated session.

Benchmark Performance

Claude Opus 4.8 across independent evaluations — coding, math, science, and reliability.

 Artificial Analysis Intelligence Index 61.4 
Independent composite score across dozens of tasks. #1 among all frontier models. Beats GPT-5.5 Pro (60.2) and Gemini 3.1 Pro (57.0).

 SWE-bench Verified (Coding) 88.6% 
Gold-standard real-world software engineering benchmark. Resolving actual GitHub issues. Highest score among frontier text models.

 USAMO 2026 (Mathematics) 96.7% 
USA Mathematical Olympiad. From 69.3% on Opus 4.7 to 96.7% — a 27-point gain. #1 on this benchmark.

Online-Mind2Web (Browser Agent) 84%

Multi-step web task completion. Highest among frontier models. 100% on Anthropic’s Super-Agent internal eval.

AA-Omniscience (Hallucination Rate) 35.9%

Lower is better. Claude Opus 4.8 achieves the lowest hallucination rate among frontier models — substantially below Google and OpenAI peers.

Humanity’s Last Exam (HLE) #1

Expert-level questions across all scientific disciplines. Claude Opus 4.8 leads the HLE leaderboard. Also #1 on CritPt frontier physics eval.

What Claude Opus 4.8 Is Built For

Agentic Coding & Refactoring

Spawn parallel subagents to migrate codebases, write test suites, refactor architecture, and fix bugs across hundreds of thousands of lines. Claude Code runs Opus 4.8 natively.

Prompt

“Refactor this Express.js API to TypeScript, add input validation with Zod, generate Jest tests for every endpoint, and open a PR with a summary of all changes made.”

Scientific Research & Analysis

GPQA Diamond, HLE, CritPt — Claude Opus 4.8 leads on every frontier science benchmark. Ideal for literature synthesis, hypothesis generation, and quantitative analysis.

Prompt

“Synthesize the last 50 papers on mRNA vaccine delivery mechanisms. Identify convergent findings, contradictions, and the three most promising unexplored research directions.”

Enterprise Document Processing

1M token context handles entire contract libraries, audit trails, and policy corpora. Extract, compare, classify, and summarize at scale with the lowest hallucination rate in the field.

Prompt

“Review these 47 vendor contracts. Flag any non-standard indemnification clauses, identify contracts expiring in the next 90 days, and generate an executive risk summary.”

Browser Automation & Computer Use

84% on Online-Mind2Web. Claude Opus 4.8 navigates real websites, fills forms, interacts with desktop UIs, and completes multi-step tasks without human handholding.

Prompt

“Go to our CRM, pull all leads tagged 'enterprise' created this month, cross-reference with LinkedIn to verify company size, and output a prioritized outreach list.”

Honest Limitations

Where Claude Opus 4.8 isn’t the best choice — and what to use instead

Terminal-Bench Coding Tasks

GPT-5.5 Pro leads on Terminal-Bench 2.0 (82.7% vs Claude’s ~75%). For pure code execution benchmarks in terminal/shell environments, GPT-5.5 Pro currently has an edge. Claude Opus 4.8 leads on real-world SWE-bench, but if your workflow is Terminal-Bench specific, test both.

Price vs Budget Models

Standard pricing of $5/$25 per M tokens is competitive against GPT-5.5 Pro ($5/$30) but significantly higher than DeepSeek V4 Pro ($0.435/$0.87). For high-volume, cost-sensitive tasks where response quality is secondary, DeepSeek or Gemini Flash may be more economical.

Microsoft Foundry Context Limit

On Microsoft Foundry (Azure), Claude Opus 4.8 is limited to 200K tokens context rather than the full 1M available on Anthropic API, Bedrock, and Vertex AI. If your workflow requires maximum context on Azure, plan accordingly.

Multimodal Video Understanding

Claude Opus 4.8 excels at text, code, images, and documents. For native video understanding and analysis, Gemini 3.1 Pro (with its Antigravity architecture) has a stronger native video pipeline. Claude handles images from video frames but isn’t a dedicated video model.

Real-Time Web Access

Claude Opus 4.8 does not have built-in real-time internet access in standard API mode. For tasks requiring current events, live pricing, or up-to-the-minute data, use Claude with tool use and a search API, or consider models with native search integration.

What Experts Say About Claude Opus 4.8

Topped the Intelligence Index

61.4 — the highest independent composite score of any frontier model.

Claude Opus 4.8 is the new leader on the Artificial Analysis Intelligence Index. Opus 4.8 scores 61.4, up +4.1 points from Opus 4.7 and +1.2 points ahead of GPT-5.5, the previous Index leader.

— Artificial Analysis, May 2026

Dramatically More Honest

4× less likely to let code flaws pass unmentioned versus its predecessor.

Opus 4.8 is around four times less likely than its predecessor to allow flaws in code it has written to pass unremarked. It achieved the lowest incorrect-rate of six models on every benchmark — the most direct measure of factual hallucination.

— Simon Willison, simonwillison.net, May 28 2026

SWE-bench Leader

Highest real-world coding score among frontier text models.

Opus 4.8 improves across benchmarks, especially the agentic-work stack. SWE-Bench Verified: 88.6%. The headline upgrade is judgment, not raw intelligence: Anthropic says it is around four times less likely than Opus 4.7 to let flaws in its own code pass unremarked.

— Karo Zieminski, Product with Attitude, Substack, 2026

Frequently Asked Questions

Claude Opus 4.8 is Anthropic’s current flagship model, released in May 2026. Key upgrades over Opus 4.7: Adaptive Thinking (configurable effort levels), Dynamic Workflows (parallel subagent orchestration), Fast Mode (2.5× throughput at lower cost), mid-conversation system messages, 4× improved honesty, a 27-point USAMO math gain, and the #1 position on the Artificial Analysis Intelligence Index. The API is fully backward-compatible — no breaking changes.

Adaptive Thinking lets Claude automatically scale its reasoning depth based on task complexity. You can also set it manually via five effort levels: low, medium, high, xhigh, and max. Low effort gives fast, token-efficient answers for simple queries. Max effort triggers full extended thinking for complex proofs, architectural decisions, and multi-step reasoning. This eliminates wasted compute on trivial tasks while unlocking maximum capability for hard problems.

Dynamic Workflows let Claude Opus 4.8 act as an orchestrator: it receives a high-level objective, breaks it into subtasks, spawns hundreds of parallel Claude subagents to handle them simultaneously, verifies each result, and synthesizes a final output. This is how codebase-scale migrations, multi-source research synthesis, and batch document processing run autonomously. Available through Claude Code and the Anthropic API.

Fast Mode is a high-throughput variant of Claude Opus 4.8 optimized for latency-sensitive and cost-sensitive workloads. Pricing: $10 per million input tokens and $50 per million output tokens (standard is $15/$75 on some configurations). Throughput increases 2.5×. Best for agentic loops with many short calls, real-time applications, and high-volume document processing where you don’t need maximum reasoning depth.

1 million tokens on Claude API, Amazon Bedrock, and Google Vertex AI. 200K tokens on Microsoft Foundry (Azure). Maximum output per call: 128K tokens. The 1M context handles entire codebases, multi-year conversation histories, full legal contracts, and large research datasets in a single request without chunking or retrieval workarounds.

Claude Opus 4.8 leads on: Intelligence Index (61.4 vs 60.2), SWE-bench Verified (88.6% vs ~80%), USAMO math (96.7%), hallucination rate (35.9% vs ~45%), browser agent (84% vs 78.7%), and output pricing ($25 vs $30/M). GPT-5.5 Pro leads on: Terminal-Bench 2.0 coding tasks (82.7%). Both have 1M token context at the same $5/M input price. For most production workloads, Claude Opus 4.8 benchmarks higher across the most comprehensive independent evaluation.

Anthropic made a deliberate architectural and training decision to prioritize honesty over helpfulness when the two conflict. Claude Opus 4.8 is explicitly trained to abstain on uncertain questions rather than generate plausible-sounding wrong answers. The sycophancy-under-pressure eval score is 0%: the model doesn’t flip correct answers when a user pushes back. This reflects Anthropic’s Constitutional AI approach to model alignment.

A new Claude Opus 4.8 API feature that lets you update the system prompt mid-session without starting a new conversation. In practice: you can change the model’s role, language, constraints, or persona in the middle of an ongoing dialogue. Useful for multi-step agentic workflows where different phases require different system instructions, or for switching between user-facing and tool-calling modes.

Yes. Claude Opus 4.8 is available on Anthropic’s own API, Amazon Bedrock, and Google Vertex AI — all with the full 1M token context window. On Microsoft Foundry (Azure), it’s available with a 200K context limit. Prompt caching is supported across all major providers. Batch API (50% price reduction) is available on the Anthropic API.

Standard: $5.00 input / $25.00 output per million tokens. Fast Mode: $10/$50 per M tokens (2.5× throughput). Batch API: $2.50/$12.50 per M tokens (50% discount, async delivery). Prompt cache write: $3.75/M, cache read: $0.30/M. Pricing is unchanged from Claude Opus 4.7 at the standard tier, so existing applications have predictable cost profiles.