Claude
Claude Opus 5 & Fable 5 — Anthropic
Anthropic's precision agent for complex tasks. Opus 5 is the new default everywhere — 96.0% on SWE-bench Verified, thinking on by default, and a habit of verifying its own work before claiming it's done. Fable 5 stays the ceiling for autonomous runs that last days. The 1M context still reads your entire codebase and remembers page 847.
Opus 5 Released
Jul 24, 2026
Context
1M tokens
Starting Price
$20/mo (Pro)
Max Output
128K tokens
01
Overview
Claude is Anthropic's flagship AI assistant, designed with a focus on being helpful, harmless, and honest. It's not the flashiest name in AI, but it's the model that developers and enterprises keep coming back to when the work actually matters.
Anthropic shipped four models in under two months, and Claude Opus 5 (July 24, 2026) is the one that matters for most people. It's now the default across claude.ai, Claude Code, and Cowork — and it landed at the same $5 / $25 per million tokens Opus 4.8 charged, while roughly doubling agentic-coding performance.
The headline change is behavioral, not spec-sheet. Opus 5 verifies its own work and iterates until it actually succeeds, instead of announcing victory after one clean run. That's the single biggest quality-of-life shift in a coding agent since parallel subagents — and it means you should delete the "double-check your work" instructions you added for older models. They now cause over-verification.
The numbers back it up: 96.0% on SWE-bench Verified (effectively saturating the benchmark), 79.2% on SWE-bench Pro, 43.3% on Frontier-Bench agentic coding — ahead of both Fable 5 (33.7%) and GPT-5.6 Sol (34.4%) — and 30.2% on ARC-AGI-3, roughly three times the next-best model. On GDPval-AA v2 knowledge work it scores 1,861 Elo vs. Fable 5's 1,747.
Claude Fable 5 is still the ceiling and still costs $10 / $50. It edges Opus 5 on SWE-bench Pro (80.0% vs 79.2%) and remains the pick for long-horizon agents that run autonomously for days. For everything else, Opus 5 matches or beats it at half the input price. Thinking is always on with Fable 5; on Opus 5 it's on by default but you can disable it at effort high or lower.
Above both sits Claude Mythos 5: same capabilities as Fable 5, available only through "Project Glasswing" to infrastructure providers and vetted cybersecurity researchers. Not a general release, and Opus 5 deliberately trails it on cyber-exploitation tasks.
Access the free tier at claude.ai (runs on Sonnet 4.6) or upgrade to Pro ($20/mo), where Opus 5 is the strongest model available. On Max it's the default. Also on the API as claude-opus-5, plus Amazon Bedrock, Google Vertex AI, GitHub Copilot, and Microsoft Foundry — Copilot picked it up on launch day.
02
Key Capabilities
Self-Verification (Opus 5)
Checks its own work and iterates until it succeeds, unprompted. Delete your "double-check your answer" instructions — on Opus 5 they cause over-verification
Benchmark Sweep (Opus 5)
96.0% SWE-bench Verified, 79.2% SWE-bench Pro, 43.3% Frontier-Bench agentic coding (vs 34.4% for GPT-5.6 Sol), 30.2% ARC-AGI-3 — about 3x the next-best model
Effort Control
Five levels — low, medium, high, xhigh, max. Same per-token rate at every level. Start at xhigh for coding, then sweep down: low and medium punch well above their weight on Opus 5.
Fast Mode (Opus 5)
2.5x faster output for double the rate ($10 / $50 per million). Claude API and Claude Code only — not on Bedrock, Vertex, or Foundry
High-Resolution Vision
High-res image support up to 2576px / 3.75MP. Big jump for design reviews, dense diagrams, and document analysis
Adaptive Thinking
Automatically adjusts reasoning depth — fast for simple tasks, deep for complex ones
Multi-Agent Coordination
Runs writer-verifier patterns cleanly and rarely lets subagents overwrite each other. Opus 5 delegates more readily than 4.8, so cap spawn counts if you're cost-sensitive
Computer Use
Most accurate model for operating computers — navigates interfaces, fills forms, moves data across applications
1M Token Context (Standard)
1M is both the default and the maximum on Opus 5 — no beta header, no separate tier. Process entire codebases, legal documents, or research papers in one conversation
128K Max Output (300K Beta)
128K standard, or up to 300K tokens via beta header (output-300k-2026-03-24)
Compaction
Automatic server-side context summarization when conversations approach the limit
Cheaper Prompt Caching
Minimum cacheable prompt drops to 512 tokens on Opus 5, down from 1024 on Opus 4.8 — prompts you'd written off as uncacheable now cache with zero code changes
Long-Horizon Autonomy (Fable 5)
Fable 5 remains the pick for agents that run unattended for days. Single requests can take many minutes — plan your timeouts and progress UI accordingly
Long-Context Coherence
76% accuracy on MRCR v2 8-needle 1M variant
03
Pricing
| Tier | Price | Details |
|---|---|---|
| Free | $0 | Runs on Sonnet 4.6. Projects, Artifacts, app connectors. Tighter limits. |
| Pro | $20/mo ($17/mo annual) | ~5x free tier messages. Opus 5 is the strongest model available on Pro. |
| Max 5x | $100/mo | 25x free tier capacity. Opus 5 is the default model. |
| Max 20x | $200/mo | 100x free tier capacity. Opus 5 default, Fable 5 for the heaviest runs. |
| Team | $25-$30/user/mo | Collaboration features, admin controls, Opus 5 access. |
| Enterprise | Custom | SSO, audit logging, compliance APIs, enhanced context. |
| API — Opus 5 | $5 / $25 per 1M tokens | The default. Fast Mode $10 / $50 at 2.5x speed. Batch API at 50% off. |
| API — Fable 5 | $10 / $50 per 1M tokens | Top tier, for multi-day autonomous agents. Requires 30-day data retention. |
| API — Sonnet 4.6 | $3 / $15 per 1M tokens | Mid-tier performance at lower cost. |
| API — Haiku 4.5 | $1 / $5 per 1M tokens | Fastest, most affordable. |
04
Best Use Cases
Complex software development and large codebase management
Long, sustained coding sessions and agentic coding tasks
Multi-application workflow automation via computer use
Enterprise document analysis — legal, financial, medical
Research tasks requiring long-context coherence across massive documents
Building AI-powered applications via the API
Tasks requiring safety, compliance, and governance
05
Pro Prompts
Codebase Analysis
Debug this large codebase and suggest refactors. Prioritize by impact-to-effort ratio. Flag any security concerns you spot.Scope Discipline (Opus 5)
Deliver exactly what I asked for, at the scope I intended. Make routine judgment calls yourself; check in only when two readings would lead to materially different work. If you think the ask is wrong, say so in one sentence and keep going. Don't add helpers, abstractions, or error handling for cases that can't happen. Task: [task].Max Effort (Fable 5)
Set effort to max. This is the kind of problem that justifies the cost: [hard problem]. Take the time. Verify every step. I'd rather wait 5 minutes for a right answer than 30 seconds for a wrong one.Document Deep Dive
Analyze this 200-page contract. Extract every obligation, deadline, and liability clause. Flag anything that deviates from standard commercial terms.Computer Use
Navigate to [application], export the Q4 report as CSV, clean the data, and create a summary dashboard in Google Sheets.06
Limitations
Higher latency at high/max effort levels, especially on complex reasoning tasks
Opus 5 thinks by default, and max_tokens caps thinking plus response together — budgets tuned for older models can truncate mid-answer
Spatial reasoning remains a common weakness
Opus 5 reaches for subagents more eagerly than 4.8 did — great for coverage, expensive if you don't cap spawn counts
Free tier does NOT include Opus 5 or Fable 5 — runs on Sonnet 4.6
Fable 5 is 2x the per-token cost of Opus 5 ($10/$50 vs $5/$25), and requires 30-day data retention — zero-retention orgs get a 400 on every request
Both Opus 5 and Fable 5 ship cybersecurity and biology safeguards that can refuse a request outright. Legitimate security and life-sciences work does hit false positives.
Claude Mythos 5 (the unrestricted tier) is gated behind Project Glasswing — infrastructure providers and vetted cybersecurity researchers only
Writes longer answers and longer files than previous models. Lowering effort doesn't reliably shorten them — you have to ask for concision explicitly.
No native image generation — text and code output only
07
Multimodal Capabilities
Vision (Up to 3.75MP)
- High-res charts, graphs, technical diagrams
- Reports, screenshots, dense documents
- Max image input: 2576px / 3.75MP
Computer Use
- Desktop software operation
- GUI navigation and form filling
File Uploads
- PDFs and documents
- Spreadsheets and presentations
Output
- Text and code only
- No native image, audio, or video generation
08
Recent Updates
Aug 3, 2026
Claude Code v2.1.221 — Focus view collapses tool activity into compact summaries and declutters long sessions
Jul 24, 2026
Claude Opus 5 generally available — 96.0% SWE-bench Verified, self-verification, 1M context standard, $5/$25 pricing. Default across claude.ai, Claude Code, and Cowork, and live in GitHub Copilot the same day.
Jun 9, 2026
Claude Fable 5 generally available — the top tier, $10/$50 API pricing. Mythos 5 deployed the same day via "Project Glasswing" for infrastructure providers and vetted cybersecurity researchers.
May 28, 2026
Opus 4.8 launched — Dynamic Workflows (parallel subagents), effort control, cheaper Fast Mode. Superseded by Opus 5 at identical pricing.
Apr 16, 2026
Claude Opus 4.7 generally available — step-change agentic coding, high-res vision (3.75MP)
Milestone
Available across all Claude products and the API, plus Amazon Bedrock, Google Vertex AI, GitHub Copilot, and Microsoft Foundry