Anthropic has released Claude Mythos 5 — a roughly 10-trillion parameter frontier model that the company says shatters state-of-the-art benchmarks across reasoning, coding, and multi-modal understanding. The release lands as Anthropic’s annualised revenue is reported to have crossed US$30 billion — a run-rate that has doubled in roughly two months.
Mythos 5 is available today through the Anthropic API, the Claude consumer product, and partner surfaces including Amazon Bedrock and Google Cloud Vertex AI.
Benchmark Claims
Anthropic’s model card reports material gains over the prior Claude 4.7 generation:
- SWE-Bench Verified: 91.8% (vs ~76% previous frontier)
- GPQA Diamond: 87.4% single-shot
- MMLU-Pro: 89.2%
- AIME 2025: 95.6% with thinking budget
- Agentic workbench gains of 30–45% on long-horizon tool-use tasks vs Claude 4.7
Independent re-runs typically land 2–4 points below vendor numbers, but even after haircuts Mythos 5 appears to clear GPT-5.4 Pro and Gemini 3.1 Ultra on most third-party replications.
What Actually Changed
The technical note is unusually candid:
- ~10T parameters via a deep mixture-of-experts architecture with sparse routing
- Longer effective context — up to 2 million tokens with materially improved retrieval-within-context
- Thinking blocks redesign — per-turn thinking budget is tunable and chargeable separately
- Agentic memory — persistent, portable memory primitives wired directly into the Messages API
- Tool-use trainer — a new curriculum around long-running, multi-step agent trajectories
Business Context
Anthropic’s annualised run-rate has climbed from roughly US$14 billion in February to about US$30 billion this week — the steepest ARR ramp in enterprise software history. Coding workloads drive the majority of the mix; Anthropic’s share of the AI coding market is estimated at above 40%. The company raised a US$30 billion Series G earlier this month at a valuation near US$400 billion.
Pricing
Mythos 5 is priced at parity with the top SKU from the prior generation — a deliberate move to avoid ceding enterprise spend to OpenAI and Google at a moment when Anthropic still has room to grow. Prompt caching pricing is unchanged. Batch API remains 50% discount.
What to Watch
Three signals over the next 30 days will determine whether Mythos 5 holds its lead:
- OpenAI’s response — GPT-5.4 Pro is already in market; a Pro+ or GPT-6 preview could reset benchmarks
- Gemini 4 — Google I/O is 19 May 2026; a frontier jump there would pull enterprises back into triple evaluations
- Real-world agentic wins — Mythos’s benchmark gains matter only if they translate into measurable enterprise ROI at scale
Source: Anthropic Newsroom / The Information
Leave a Reply