AnthropicProprietary

Claude Opus 4.8

Name: Claude Opus 4.8
Author: Anthropic

The latest foundation model developed by Anthropic. It achieves significantly improved performance and efficiency compared to the previous generation.

Share:X Hatena LINE

Parameters

Undisclosed

Context Window

License

Proprietary

Release Date

2026-05-28

Japanese Language Capability

✅High-Quality JP

Multilingual model with strong Japanese language processing capabilities.

API Pricing

Input Price (per 1M tokens)

Output Price (per 1M tokens)

Billing Mode: standard

Strengths

Weaknesses

Use Cases

Deep Analysis

Arena Elo

1477

BenchLM overall; 1537 on coding subset

SWE-bench Pro

69.2%

Field-leading; GPT-5.5: 58.6%, Gemini 3.1 Pro: 54.2%

SWE-bench Verified

88.6%

Up from 87.6% on Opus 4.7

GDPval-AA

1,890 Elo

67% implied win rate vs GPT-5.5

Intelligence Index

61.4

Was #1 at release; +4.1 over Opus 4.7

Input / Output Price

$5 / $25 per 1M

Unchanged from Opus 4.7; fast mode $10/$50

Strengths

・Field-leading agentic coding performance (SWE-bench Pro 69.2%) with significant margin over GPT-5.5 and Gemini 3.1 Pro
・~4x reduction in letting code flaws pass without flagging, improving reliability in long-horizon agentic workflows
・Strongest long-context reasoning at 1M tokens (GraphWalks BFS 68.1% vs GPT-5.5's 45.4%) and leading knowledge-work Elo (GDPval-AA 1,890)

Weaknesses

・GPT-5.5 still leads on terminal-based DevOps coding (Terminal-Bench 78.2% vs 74.6%) and uses ~30% fewer turns for agentic tasks
・Prompt injection resistance regressed from ~2.3% to ~7% attack success rate, a concern for adversarial environments
・High token consumption and output pricing ($25/M) makes it expensive vs Gemini 3.1 Pro ($12/M output) for cost-sensitive high-volume workloads

Competitor Comparison

Model	Arena	GPQA	Price
GPT-5.5 (xhigh)	~1480*	~92%*	$25/$30
Gemini 3.1 Pro	~1460*	~92%*	$2/$12
Claude Fable 5	N/A	N/A	$10/$50

Overview

Claude Opus 4.8, released May 28, 2026, is Anthropic's incremental but substantive upgrade to the Opus 4 line. It builds on Opus 4.7 with meaningful gains in agentic coding (SWE-bench Pro from 64.3% to 69.2%), knowledge work (GDPval-AA from 1,753 to 1,890 Elo), and scientific reasoning (Humanity's Last Exam up ~3 points), while maintaining the same $5/$25 pricing and 1M-token context window. At launch it claimed the #1 spot on the Artificial Analysis Intelligence Index at 61.4, leapfrogging GPT-5.5 by 1.2 points. Beyond benchmark numbers, the release's most consequential improvements are behavioral: Opus 4.8 is approximately four times less likely than its predecessor to let code flaws pass unremarked, and early testers consistently report better judgment, more appropriate pushback on flawed plans, and stronger self-correction in agentic workflows. The model also introduces Dynamic Workflows (a research preview enabling up to 1,000 parallel subagents in Claude Code), Effort Control (a 5-level dial for reasoning depth), and a new Fast Mode at $10/$50 offering 2.5x throughput. Anthropic positioned Opus 4.8 as the dependable production tier beneath the newly announced Fable 5 and Mythos-class models, which handle even more demanding tasks at higher price points. The model's primary weaknesses are in areas where competitors retain structural advantages: GPT-5.5 leads on terminal-based coding tasks and achieves higher turn efficiency on agentic benchmarks, Gemini 3.1 Pro is roughly 2x cheaper on output tokens for cost-sensitive workloads, and prompt injection resistance regressed from the prior version. Creative writing remains a lateral move rather than an improvement. Nevertheless, for teams building agentic coding pipelines, code review automation, or long-context analysis tools, Opus 4.8 represents the strongest mainstream-tier option as of mid-2026.

Benchmarks & Performance

## Comprehensive Benchmark Comparison ### Coding & Software Engineering | Benchmark | Claude Opus 4.8 | Claude Opus 4.7 | GPT-5.5 | Gemini 3.1 Pro | |---|---|---|---|---| | SWE-bench Verified | 88.6% | 87.6% | — | 80.6% | | SWE-bench Pro | 69.2% | 64.3% | 58.6% | 54.2% | | Terminal-Bench 2.1 | 74.6% | 66.1% | 78.2% | 70.3% | | OSWorld-Verified | 83.4% | 82.8% | 78.7% | 76.2% | | SWE-bench Multilingual | 84.4% | — | — | — | | SWE-bench Multimodal | 38.4% | — | — | — | ### Reasoning & Knowledge | Benchmark | Claude Opus 4.8 | Claude Opus 4.7 | GPT-5.5 | Gemini 3.1 Pro | |---|---|---|---|---| | GPQA Diamond | 93.6% | 91.4% | ~92% | ~92% | | Humanity's Last Exam (tools) | 57.9% | 54.7% | 52.2% | 51.4% | | Humanity's Last Exam (no tools) | 49.8% | 46.9% | — | — | | CritPt (physics) | 20.9% | — | higher | higher | | Intelligence Index (AA) | 61.4 | 57.3 | 60.2 | — | ### Agentic & Knowledge Work | Benchmark | Claude Opus 4.8 | Claude Opus 4.7 | GPT-5.5 | Gemini 3.1 Pro | |---|---|---|---|---| | GDPval-AA (Elo) | 1,890 | 1,753 | 1,769 | 1,314 | | AA-Omniscience Index | 27.4 | ~26.5 | — | 32.9 (Pro leads) | | Hallucination Rate | 35.9% | ~36% | higher | higher | | GraphWalks BFS (1M) | 68.1% | — | 45.4% | — | | GraphWalks BFS (256K) | 85.9% | — | — | — | | BrowseComp | 84.3% | — | — | — | | Finance Agent v2 | 53.9% | 51.5% | — | 57.9% (Flash) | ### Alignment & Safety | Metric | Claude Opus 4.8 | Claude Opus 4.7 | |---|---|---| | Code summary dishonesty rate | 3.7% | ~15% (est. 4x worse) | | Misalignment behavior score | 1.9 | 2.5 | | Prompt injection attack success | ~7% | ~2.3% | | Prosocial trait scores | New highs | Previous highs | ### Arena Performance (BenchLM) | Category | Elo | Rank | |---|---|---| | Overall | 1,477 | Top 10 | | Coding | 1,537 | Top 5 | | Hard Prompts | 1,510 | Top 5 | | Multi-turn | 1,496 | Top 10 | | Instruction Following | 1,481 | — | | Creative Writing | 1,461 | — |

Detailed Comparison

## Head-to-Head Comparisons ### Claude Opus 4.8 vs GPT-5.5 (OpenAI) | Dimension | Claude Opus 4.8 | GPT-5.5 | |---|---|---| | Pricing (input/output) | $5/$25 | $25/$30 | | Context window | 1M tokens | 1M tokens | | SWE-bench Pro | 69.2% | 58.6% | | Terminal-Bench 2.1 | 74.6% | 78.2% | | GDPval-AA | 1,890 Elo | 1,769 Elo | | Intelligence Index | 61.4 | 60.2 | | Turn efficiency | ~30% more turns | Baseline | | GraphWalks BFS (1M) | 68.1% | 45.4% | **Verdict:** Opus 4.8 is significantly cheaper per token and leads on agentic coding (SWE-bench Pro by 10.6 points) and knowledge work (GDPval-AA by 121 Elo). GPT-5.5 retains the lead on terminal-based DevOps tasks and is more turn-efficient, which can offset its higher per-token cost on some workloads. For long-context reasoning at 1M tokens, Opus 4.8 is substantially ahead. ### Claude Opus 4.8 vs Gemini 3.1 Pro (Google DeepMind) | Dimension | Claude Opus 4.8 | Gemini 3.1 Pro | |---|---|---| | Pricing (input/output) | $5/$25 | $2/$12 | | Context window | 1M tokens | 1M+ tokens | | SWE-bench Pro | 69.2% | 54.2% | | Terminal-Bench 2.1 | 74.6% | 70.3% | | GDPval-AA | 1,890 Elo | 1,314 Elo | | AA-Omniscience | 27.4 | 32.9 | | Hallucination rate | 35.9% | higher | **Verdict:** Gemini 3.1 Pro is roughly 2x cheaper on output tokens and leads on AA-Omniscience factuality. Opus 4.8 dominates on coding accuracy (SWE-bench Pro by 15 points) and knowledge work (GDPval-AA by 576 Elo). For teams where cost is the primary constraint and coding accuracy margins don't justify the price multiple, Gemini is the pragmatic choice. For agentic coding and professional knowledge work, Opus 4.8 delivers meaningfully better results. ### Claude Opus 4.8 vs Claude Fable 5 (Anthropic) | Dimension | Claude Opus 4.8 | Claude Fable 5 | |---|---|---| | Pricing (input/output) | $5/$25 | $10/$50 | | Context window | 1M tokens | 1M tokens | | SWE-bench Pro | 69.2% | 80.4% | | SWE-bench Verified | 88.6% | 95.0% | | GDPval-AA | 1,890 Elo | 1,932 Elo | | Retention requirement | None | 30-day retention | | Safety fallback | — | Falls back to Opus 4.8 | **Verdict:** Fable 5 is Anthropic's new top tier, beating Opus 4.8 decisively on coding benchmarks (SWE-bench Pro +11.2 points) at 2x the price with a 30-day data retention requirement. Opus 4.8 serves as both the mainstream production model and Fable 5's safety fallback for flagged sessions, making it the dependable workhorse for teams not ready for Fable's terms or cost.

Community Feedback

## Community Reception **Developer and analyst consensus:** Claude Opus 4.8 has been received as a meaningful but not revolutionary upgrade. The consensus across reviews is that it excels at what Claude was already strong in (coding, structured analysis, agentic workflows) while showing less improvement in creative writing and some reasoning tasks. **Coding community (positive):** Cursor reported that "Opus 4.8 exceeds prior Opus models across every effort level" and that "tool calling is meaningfully more efficient, using fewer steps for the same intelligence." Devin (Cognition) noted it "uses tools cleanly and follows instructions with the consistency our autonomous engineering workloads need." The SWE-bench Pro lead of 10.6 points over GPT-5.5 generated significant attention among developer-focused communities. **Enterprise users (positive):** Databricks reported a "step change in agentic reasoning" in their Genie AI agent. Bridgewater Associates praised the model's tendency to "proactively flag issues with inputs and outputs of an analysis." Harvey Legal Benchmark showed the model setting new highs for legal AI accuracy. **Honesty theme (notable):** Multiple reviewers highlighted the improved honesty and self-correction. The Android Authority review found Opus 4.8 is "not afraid to share its thoughts" and flags impossible premises rather than plowing ahead. The ~4x reduction in code flaw pass-through was cited by Awesome Agents as potentially more valuable than benchmark gains. **Criticism (mixed):** The Tech Yahoo review found creative writing to be a "lateral move at best" and flagged that Opus 4.8 failed a reasoning whodunit that Opus 4.7 solved correctly. Token consumption remains a sore point — one review noted a single coding prompt drained an entire Pro plan token quota. The 300K needle-in-haystack test collapsed entirely, though the 85K version worked (after the model refused to report its findings due to safety training). **Reddit/community sentiment:** Sentiment is broadly positive among coding-focused users who value reliability improvements, with frustration directed at token costs and the rapid release cadence (41 days between 4.7 and 4.8) that requires prompt re-tuning. The model's role as Fable 5's safety fallback has been noted as both a vote of confidence and an indication that Anthropic views Opus 4.8 as the production tier rather than the frontier.

Use Cases

## Use Case Deep Dive ### 1. Agentic Coding and Code Review Automation **When to choose Opus 4.8:** This is the model's strongest use case. With 69.2% on SWE-bench Pro (the anti-memorization benchmark), Opus 4.8 leads the field for multi-file software engineering tasks. The ~4x improvement in catching its own code flaws makes it particularly valuable for autonomous code review pipelines where a missed bug compounds through subsequent steps. Dynamic Workflows enables codebase-scale parallel work — early adopters have used it for 750,000-line migrations. Choose Opus 4.8 over GPT-5.5 when your work is repository-level (multi-file diffs, refactoring) rather than terminal-heavy DevOps. Choose over Gemini 3.1 Pro when coding accuracy margins justify the 2x output price premium. ### 2. Long-Context Analysis and Knowledge Work **When to choose Opus 4.8:** The 1M-token context window with strong reasoning at that scale (GraphWalks BFS 68.1% vs GPT-5.5's 45.4%) makes this the best model for tasks requiring analysis of large document corpora. Finance teams at Hebbia report "noticeably better citation precision" on dense filings. Databricks uses it in their Genie agent for reasoning over PDFs, diagrams, and unstructured content at 61% cheaper token cost than Opus 4.7. GDPval-AA at 1,890 Elo (67% win rate vs GPT-5.5) confirms the knowledge-work advantage. Choose this over GPT-5.5 for financial, legal, or document-heavy professional workflows where accuracy and citation quality matter. ### 3. Scientific and Academic Reasoning **When to choose Opus 4.8:** Opus 4.8 leads Humanity's Last Exam at 57.9% with tools and scores 93.6% on GPQA Diamond, closing the gap that previous Claude releases had on academic reasoning. It now slightly leads on CritPt (physics) over Gemini 3.1 Pro. For research teams needing a single model to handle scientific paper analysis, experimental design, and domain-specific reasoning, Opus 4.8 offers the best balance of reasoning quality and cost at the mainstream tier. ### 4. Multi-Step Agent Orchestration with Reliability Requirements **When to choose Opus 4.8:** The combination of improved honesty (4x fewer silently-flawed code outputs), better judgment in agentic tasks (catching mistakes, pushing back on flawed plans), and the Dynamic Workflows capability for parallel subagents makes this the right choice for teams building autonomous agents that need to run unattended. Legal teams at CoCounsel report "meaningful improvements in consistency and reasoning quality." The effort control API lets you tune reasoning depth per subtask — low effort for simple lookups, max effort for complex analysis — optimizing the cost-performance curve across a pipeline. Choose this when your agent runs for hours and cannot afford to silently produce flawed intermediate outputs. ### When NOT to choose Opus 4.8: - **Terminal-heavy DevOps:** GPT-5.5 leads Terminal-Bench 2.1 by 3.6 points - **Cost-constrained high-volume workloads:** Gemini 3.1 Pro is ~2x cheaper on output - **Creative writing:** Improvement over 4.7 is negligible; other models may serve better - **Adversarial environments:** Prompt injection resistance regressed to ~7% attack success - **Maximum capability at any cost:** Claude Fable 5 beats it on every coding benchmark at 2x the price

Latest News

## Latest Developments (as of mid-2026) **Release (May 28, 2026):** Claude Opus 4.8 launched with identical pricing to Opus 4.7 ($5/$25), the same 1M-token context window, and several new features including Dynamic Workflows (research preview for up to 1,000 parallel subagents), Effort Control (5-level reasoning dial), Fast Mode ($10/$50 at 2.5x speed, 3x cheaper than previous fast tiers), and mid-conversation system message support in the Messages API. **Prompt cache improvements:** Minimum context window for prompt caching dropped from 4,096 to 1,024 tokens, enabling automatic caching for shorter prompts with no code changes required. **Fable 5 relationship (June 9, 2026):** Twelve days after Opus 4.8, Anthropic released Claude Fable 5 as a new premium tier ($10/$50 with 30-day data retention). Opus 4.8 serves as Fable 5's safety-classifier fallback — when Fable's safety systems flag a session, the request is routed to Opus 4.8, making it the de facto "safe surface" of the Mythos era. **Claude Mythos Preview (announced alongside Opus 4.8):** Anthropic mentioned Project Glasswing, with a small number of organizations using Claude Mythos Preview for cybersecurity work. The Claude Opus 4.8 System Card references Mythos as "our best-aligned model," suggesting it represents a higher capability tier with stronger safety constraints. **Valuation milestone:** Anthropic closed a $65 billion Series H at a $965 billion valuation on the same day as the Opus 4.8 launch. **Ecosystem adoption:** Within days of release, Cursor, GitHub Copilot, Windsurf, CodeRabbit, and Replit adopted Opus 4.8 as default-selectable. The model is available on Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Azure AI Foundry (200K context limit on Foundry; 1M elsewhere). **Future roadmap:** Anthropic indicated plans to release models providing Opus-level capabilities at lower cost, and to bring Mythos-class models to general availability "in the coming weeks" after developing stronger cyber safeguards.

Beyond benchmark numbers, the release's most consequential improvements are behavioral: Opus 4.8 is approximately four times less likely than its predecessor to let code flaws pass unremarked, and early testers consistently report better judgment, more appropriate pushback on flawed plans, and stronger self-correction in agentic workflows. The model also introduces Dynamic Workflows (a research preview enabling up to 1,000 parallel subagents in Claude Code), Effort Control (a 5-level dial for reasoning depth), and a new Fast Mode at $10/$50 offering 2.5x throughput. Anthropic positioned Opus 4.8 as the dependable production tier beneath the newly announced Fable 5 and Mythos-class models, which handle even more demanding tasks at higher price points.

The model's primary weaknesses are in areas where competitors retain structural advantages: GPT-5.5 leads on terminal-based coding tasks and achieves higher turn efficiency on agentic benchmarks, Gemini 3.1 Pro is roughly 2x cheaper on output tokens for cost-sensitive workloads, and prompt injection resistance regressed from the prior version. Creative writing remains a lateral move rather than an improvement. Nevertheless, for teams building agentic coding pipelines, code review automation, or long-context analysis tools, Opus 4.8 represents the strongest mainstream-tier option as of mid-2026.

Sources

Analysis generated: 2026-07-17

Related Comparisons

Claude Opus 4.8vsGPT-5.5