Claude Fable 5: Anthropic's Mythos-Class Model Tops SWE-Bench

The Mythos-Class Tier: A Step Above Opus

On 9 June 2026, Anthropic released Claude Fable 5, the first publicly available model in the company's Mythos capability tier — a classification above the existing Opus family that Anthropic defines as models capable of frontier-level performance on tasks most relevant to knowledge work, scientific research, and autonomous software engineering. SWE-Bench Pro, the industry benchmark measuring a model's ability to complete real-world software engineering tasks from written issue descriptions, gives an immediate measure of that gap: Fable 5 scores 80.3 per cent, 11.1 points above Claude Opus 4.8's 69.2 per cent and 21.7 points above GPT-5.5's 58.6 per cent.

Alongside Fable 5, Anthropic also announced Claude Mythos 5. Mythos 5 is the higher-capability, restricted-access configuration of the same underlying model, intended for trusted enterprise environments running in Anthropic's Trusted Tier. Fable 5, with production safeguards built in, is the version accessible to the general developer and enterprise market through the Claude API, Amazon Bedrock, and Google Vertex AI.

Benchmark Performance: The Numbers That Matter

The full benchmark picture strengthens the case for Fable 5 as the current frontier leader on task-relevant evaluations. On SWE-Bench Verified, the companion evaluation measuring code generation quality across a curated test set of verified GitHub issues, Fable 5 scores 95.0 per cent — the highest recorded result on that test. OSWorld-Verified, which measures agentic computer use performance on real GUI tasks, stands at 85.0 per cent, above Opus 4.8's 83.4 per cent. On GDPval-AA, a research quality and knowledge synthesis benchmark, Fable 5 reaches 1932 Elo, 42 points above Opus 4.8's 1890. Humanity's Last Exam with tools, the evaluation designed to test models at the frontier of academic and research reasoning, comes in at 64.5 per cent.

What the SWE-Bench Pro Score Means in Practice

SWE-Bench Pro draws from real issues filed in production open-source repositories, requiring models to read context, identify root cause, plan a fix, write code, and pass the associated test suite — all without human intervention. A score of 80.3 per cent means Fable 5 independently resolves more than four in five issues of this type. That is not a laboratory benchmark isolated from real development work — it is close to production-grade autonomous code repair. For engineering teams evaluating agentic coding pipelines, SWE-Bench Pro is the single most honest predictor of how much autonomous throughput a model will deliver.

Context Window and Pricing

Claude Fable 5 supports a one-million-token input context window with 128,000 output tokens — sufficient for entire large codebases, full legal documents, or extended multi-session agentic workflows without truncation. Prompt caching is available at a 90 per cent input token discount, which meaningfully reduces the cost of repeated long-context calls in production systems.

Pricing is set at 10 US dollars per million input tokens and 50 US dollars per million output tokens — double the rate of Claude Opus 4.8 (5 and 25 dollars respectively). The cost premium is significant, but the benchmark gap is correspondingly large: teams choosing Fable 5 over Opus 4.8 for autonomous coding tasks get 11 additional percentage points on the most directly task-relevant benchmark available.

The June 22 Transition: What Enterprise Teams Must Plan For

Anthropic included Fable 5 in existing Pro, Max, Team, and seat-based Enterprise subscription plans at no extra cost from 9 June through 22 June 2026 — a 13-day access window. From 23 June 2026, Fable 5 is removed from flat-rate subscriptions and continued access requires purchasing usage credits billed at the API rate.

For enterprise procurement teams on annual contracts, this is an immediate action item. Usage at Fable 5's pricing inside agentic pipelines — particularly long-context tasks where the 10-dollar input rate applies to large prompts — can accumulate quickly. Teams that built their June cost model assuming subscription-included access will need to reforecast from 23 June.

What This Means for Indian Development Teams

India's AI product companies and technology services firms increasingly run agentic software engineering workloads — code review pipelines, automated test generation, documentation drafting, and multi-step planning agents — at a scale where model capability and per-token cost interact directly with project economics.

Fable 5's SWE-Bench Pro score of 80.3 per cent means substantially fewer failed autonomous code generation attempts requiring human intervention. For teams billing on outcomes, that is a real reduction in unplanned engineering hours. For teams evaluating whether to build or buy autonomous coding capabilities, Fable 5 sets a new baseline for what is achievable today.

The pricing at 10 dollars per million input tokens is above Opus 4.8 but well below what would be required to approach Fable 5's benchmark results through any alternative closed model. The June 22 subscription cut-off means organisations need to update their API cost planning now, not at month-end.

The Bottom Line

Claude Fable 5, released 9 June 2026, is the current frontier coding model: SWE-Bench Pro at 80.3 per cent, SWE-Bench Verified at 95.0 per cent, and OSWorld-Verified at 85.0 per cent, all with a one-million-token context window. Priced at 10 and 50 dollars per million tokens, it is double the cost of Opus 4.8 and produces results 11 percentage points ahead on the benchmark most relevant to autonomous software engineering. The subscription access window closes on 22 June 2026. Teams that have not yet evaluated Fable 5 for their agentic coding pipelines should do so before that date.