LIVE · SAT, JUN 20, 2026 --:--:-- ET

Issue Nº 60 COST TOTAL $14462.86 ARTICLES TODAY 12 TOKENS TOTAL 9.04B

Running the wire

Research Claude Fable 5 reaches 95% on SWE-bench Verified, tops every benchmark except GPT-5.5 Funding 201 Ventures plans second European defence fund as geopolitical tensions fuel AI and autonomy investments Funding Twenty Technologies reaches $1B valuation; defense cyber-warfare startup closes $100M Series B Policy EU AI Act transparency requirements take effect August 2, 2026 Breaking US export controls force Anthropic to suspend Claude Fable 5 for all users Policy Canada accelerates defense semiconductor strategy amid NATO spending push Research OpenAI improves health intelligence in ChatGPT; GPT-5.5 Instant reaches frontier model performance on medical evaluations Breaking CircleCI launches Chunk Sidecars: in-loop validation for AI coding agents, catching issues before commit Policy Canada invests $900M in defense industrial strategy, backing semiconductor startups and quantum tech Chips AMD reinstate memory encryption (TSME) on Ryzen 9000 CPUs via July BIOS update Chips Atom Computing secures $100M Series C plus $100M CHIPS Act backing for neutral-atom quantum scale Policy FERC issues tailored large-load interconnection orders for AI data centers; data center operators bear upgrade costs Funding Qualcomm in talks to acquire Tenstorrent for $8–10B, bet on RISC-V AI chips Breaking Amazon employees face discipline for testifying against data centers; retaliation claims filed Chips TSMC 2026 capex hits $52-56B record; 90% of world's advanced chips concentrated in one island Funding Qualcomm in talks to acquire Tenstorrent for $8–10B; RISC-V AI chip challenge to Nvidia Research MosaicLeaks: Research agents leak private data through query patterns; PA-DR training cuts leakage to 9.9% Funding Anthropic closes $65B Series H at $965B, overtakes OpenAI as most valuable AI startup Chips Trump announces Apple-Intel chip design and manufacturing partnership in US Policy Bernie Sanders files AI public ownership bill; proposes 50% stock tax, $1K annual dividends Research Claude Fable 5 reaches 95% on SWE-bench Verified, tops every benchmark except GPT-5.5 Funding 201 Ventures plans second European defence fund as geopolitical tensions fuel AI and autonomy investments Funding Twenty Technologies reaches $1B valuation; defense cyber-warfare startup closes $100M Series B Policy EU AI Act transparency requirements take effect August 2, 2026 Breaking US export controls force Anthropic to suspend Claude Fable 5 for all users Policy Canada accelerates defense semiconductor strategy amid NATO spending push Research OpenAI improves health intelligence in ChatGPT; GPT-5.5 Instant reaches frontier model performance on medical evaluations Breaking CircleCI launches Chunk Sidecars: in-loop validation for AI coding agents, catching issues before commit Policy Canada invests $900M in defense industrial strategy, backing semiconductor startups and quantum tech Chips AMD reinstate memory encryption (TSME) on Ryzen 9000 CPUs via July BIOS update Chips Atom Computing secures $100M Series C plus $100M CHIPS Act backing for neutral-atom quantum scale Policy FERC issues tailored large-load interconnection orders for AI data centers; data center operators bear upgrade costs Funding Qualcomm in talks to acquire Tenstorrent for $8–10B, bet on RISC-V AI chips Breaking Amazon employees face discipline for testifying against data centers; retaliation claims filed Chips TSMC 2026 capex hits $52-56B record; 90% of world's advanced chips concentrated in one island Funding Qualcomm in talks to acquire Tenstorrent for $8–10B; RISC-V AI chip challenge to Nvidia Research MosaicLeaks: Research agents leak private data through query patterns; PA-DR training cuts leakage to 9.9% Funding Anthropic closes $65B Series H at $965B, overtakes OpenAI as most valuable AI startup Chips Trump announces Apple-Intel chip design and manufacturing partnership in US Policy Bernie Sanders files AI public ownership bill; proposes 50% stock tax, $1K annual dividends

Research Saturday, June 20, 2026 at 06:35 AM

Claude Fable 5 reaches 95% on SWE-bench Verified, tops every benchmark except GPT-5.5

Anthropic released Claude Fable 5 on June 9, 2026, the first publicly available model from its Mythos-class tier (historically restricted to cyber-defense and biology partners). Fable 5 achieves 95.0% on SWE-bench Verified, the hardest standardized benchmark for real-world code generation, and leads or ties on 18 of 19 published benchmarks. Only GPT-5.5 beats it on a handful of reasoning and domain-specific tasks. This is the first time Anthropic has released a Mythos-class model for general use.

Fable 5 is the same underlying model as Claude Mythos 5 (restricted release) with safeguards in place: the model silently defers to Claude Opus 4.8 for requests touching cybersecurity, biology, chemistry, or model distillation, where Mythos 5's unrestricted capabilities (78% on cybersecurity evals vs. 40% for Opus 4.8) pose misuse risk. This safety-by-fallback design allows deployment at scale while maintaining alignment boundaries. Fable 5 also shows improved token efficiency: solving the same problems with fewer tokens than prior Claude generations, which compounds cost advantages at scale.

SWE-bench Verified's climb from 33.4% (Claude 3.5 Sonnet, June 2024) to 95.0% (Fable 5, June 2026) in two years reflects both model capability and benchmark saturation: the public Verified set now has a known training-data contamination history. Scale AI's SWE-bench Pro (1,865 tasks across commercial codebases) is the harder, more defensible benchmark; Fable 5 leads with 80.3% on the public set, 11 points ahead of the nearest competitor (GPT-5.5).

For architects: Fable 5's cost is 2x the Opus tier on Claude.ai; per-token pricing (~$20/M input, $60/M output) is frontier-grade. The real value is in long-horizon autonomous coding: Stripe's testing reported Fable 5 completing a 50-million-line codebase migration in a day. Teams building agentic coding pipelines should test Fable 5 on their specific codebases (Verified saturation is real); SWE-bench Pro is the more credible differentiator. The safeguards on cybersecurity mean production security-fix agents will silently degrade to Opus 4.8 for certain tasks.

Sources