Claude Opus 4.8
Anthropic's reliability-first frontier model — 88.6% on SWE-bench Verified (the highest published score), Intelligence Index 61.4, and a 1M-token context at standard pricing. Released May 28, 2026; now the top Opus-tier model below Claude Fable 5.
Model Specs
- Released
- May 2026
- Context window
- 1.0M tokens
- Max output
- 128K tokens
- Capabilities
- reasoningmultimodalfunction-callinglong-context
- Modalities
- textvision
About this model
Claude Opus 4.8 launched May 28, 2026 as Anthropic's flagship and held the #1 spot on the Artificial Analysis Intelligence Index (61.4) until Claude Fable 5 arrived twelve days later. It remains the top Opus-tier model — and arguably the best value in Anthropic's premium lineup at $5/M input and $25/M output, a fraction of what Opus-class pricing used to cost ($18.75/$75 for Opus 4.1, which is now deprecated with Opus 4.8 as its official migration target).
The launch theme was **reliability and honesty**. On AA-Omniscience, Opus 4.8 posts one of the lowest incorrect-answer behaviors among frontier models — it abstains rather than hallucinates — and Anthropic reports it is roughly 4x less likely than Opus 4.7 to leave flaws in its own code unremarked. The engineering numbers lead the industry: **88.6% on SWE-bench Verified** (the highest published score on the benchmark), 69.2% on the harder SWE-Bench Pro, 74.6% on Terminal-Bench 2.1, 82.2% on MCP-Atlas, and 93.6% on GPQA Diamond. On GDPval-AA (real-occupation tasks) it scores 1890 Elo — achieved with 15% fewer conversation turns and 35% fewer output tokens than Opus 4.7, meaning it gets to the answer more directly.
Like all post-4.7 Claude models, Opus 4.8 uses always-on adaptive thinking with an effort parameter (sampling controls are removed), accepts text and images, and offers a 1M-token context window at standard pricing with no long-context premium — plus a 90% prompt-caching discount ($0.50/M cache hits). Reach for Opus 4.8 when you want frontier engineering quality with trustworthy, low-hallucination behavior at half the price of Fable 5 — and step up to Fable 5 only for the absolute hardest problems.
Key Strengths
Highest SWE-bench Verified score published
88.6% on SWE-bench Verified — ahead of every other published result — plus 69.2% on the harder SWE-Bench Pro and 74.6% on Terminal-Bench 2.1. Real-codebase engineering is the model's core strength.
Reliability and honesty by design
One of the lowest incorrect-answer rates among frontier models on AA-Omniscience: it abstains instead of hallucinating, and it's ~4x less likely than Opus 4.7 to leave flaws in its own code unremarked.
1M context at standard pricing
Full million-token context with no long-context price premium — unlike rivals that double rates above 200K. Whole codebases and document sets at flat $5/$25 rates.
Frontier intelligence, mid-frontier price
Intelligence Index 61.4 — #1 at launch, still #2 behind only Fable 5 — at $5/$25, half of Fable 5's rate and a fraction of legacy Opus pricing.
Token-efficient agentic work
GDPval-AA 1890 Elo achieved with 15% fewer turns and 35% fewer output tokens than Opus 4.7 — it reaches answers more directly, which compounds into real cost savings on agent loops.
90% prompt-caching discount
Cache hits at $0.50/M against the $5/M base input rate. Long-running agents and repeated-prefix workflows get dramatic effective discounts.
How it compares
Opus 4.8 sits just below Fable 5 at half the price — the value point of Anthropic's frontier lineup.
| vs. Model | Verdict | Outcome |
|---|---|---|
| Claude Fable 5 | Fable 5 is a tier above (Intelligence 64.9 vs 61.4, SWE-Bench Pro 80.3% vs 69.2%, FrontierCode 29.3% vs 13.4%) but costs exactly double ($10/$50 vs $5/$25). Opus 4.8 actually leads on SWE-bench Verified (88.6%). Use Opus 4.8 as the frontier default; escalate the hardest problems to Fable 5. | Other wins |
| GPT-5.5 | Nearly tied on intelligence (61.4 vs 60.2) at comparable prices ($5/$25 vs $5/$30). Opus 4.8 leads on real-codebase engineering (SWE-Bench Pro 69.2% vs 58.6%, GDPval-AA 1890 vs 1769 Elo) and hallucinates far less; GPT-5.5 counters with computer use and the OpenAI agentic stack. For code and trust, Opus 4.8. | Wins most cases |
| Claude Sonnet 4.5 | Sonnet 4.5 is the everyday workhorse — cheaper and faster to first token. Opus 4.8 justifies its premium with a different class of engineering performance (SWE-bench Verified 88.6% vs 77.2%) and 1M context vs 200K. Route by task difficulty. | Wins most cases |
Pros
- SWE-bench Verified 88.6% — highest published score
- Intelligence Index 61.4 (#2 overall, behind only Fable 5)
- Reliability-tuned: abstains instead of hallucinating
- 1M-token context at flat pricing — no long-context premium
- Half the price of Fable 5 ($5/$25 vs $10/$50)
- Token-efficient: fewer turns and fewer output tokens than Opus 4.7
- 90% prompt-caching discount ($0.50/M cache hits)
Things to consider
- Superseded as absolute flagship by Claude Fable 5 within two weeks
- Slow first token (~57s) due to always-on adaptive thinking
- Trails Fable 5 significantly on the hardest coding tiers (FrontierCode 13.4% vs 29.3%)
- 200K context cap on Microsoft Foundry deployments
- Sampling controls (temperature/top_p/top_k) removed
- Still premium-priced versus Sonnet 4.5 for routine tasks
Best use cases
Production code engineering
The SWE-bench Verified record holder: multi-file fixes, refactors, and reviews where correctness and self-honesty about edge cases matter.
High-trust analysis
Low-hallucination behavior makes it the pick for legal, financial, and medical-adjacent work where a wrong answer is worse than no answer.
Long-horizon agents
Token-efficient turns plus 1M context plus MCP-Atlas 82.2% — agents that run longer on the same budget without losing the thread.
Full-repository comprehension
Load an entire codebase at flat pricing and ask architecture-level questions — no long-context surcharge above 200K like Gemini's tiering.
Technical writing & review
Specifications, design docs, and postmortems where precision and honest caveats beat speed.
Cost-controlled frontier workloads
Near-Fable intelligence at half the price — the rational default for teams that escalate only the hardest 10% of tasks to Fable 5.
How to use it on Renas AI
- 1
Step 1
Open AI Chat on Renas
Navigate to AI Chat in the Renas dashboard and pick the most capable Claude model available to your plan.
- 2
Step 2
Give it real engineering context
Paste whole files, error logs, and test output. Opus 4.8's edge is real-codebase work — the more genuine context, the more the SWE-bench-level quality shows.
- 3
Step 3
Trust the abstentions
When Opus 4.8 says it isn't sure, that's the reliability tuning working. Ask it to verify or provide sources rather than pushing it to guess.
- 4
Step 4
Iterate efficiently
The model reaches answers in fewer turns than its predecessors — front-load your constraints and let it produce complete solutions instead of fragmenting the task.
Pricing
Pricing on Renas AI
Pay-as-you-go credits, no API keys, no rate limits.
~66,667 words on a 10,000-credit Spark plan
Frequently asked questions
Other Anthropic models
Other text models on Renas AI
Frontier engineering on Renas
Access leading Claude models with your Renas AI subscription credits — no API key, no setup, no per-seat fees.
Try Claude on Renas