The upgrade cycle found its price tag. Anthropic’s Fable 5.1 release scored 55.8% on Terminal-Bench 4.0, up from Fable 5’s 42.0% and above Opus 5’s 52.3%, Anthropic says. Base rates stay at $10 per million input tokens and $50 per million output tokens, while cache reads fall 75% to $0.25. Gradium’s new TTS model pairs an 81.0% hard-case pass rate with 216-millisecond median first audio. Capability is improving, but production economics are doing equal work.
That cost story now comes with stricter gates. OpenAI’s Astra assessment reports 100% on ExploitBench, testing against 20 recent high-severity V8 vulnerabilities, and two zero-days used in an exploit chain. OpenAI says its cyber jailbreak refusal rate reached 91.5%, versus 59% for GPT-5.6 Sol. Anthropic says Fable 5.1’s cyber safeguards trigger about 60% fewer interventions and its biology safeguards fire 85% less often on benign requests. The frontier is splitting into broadly available capability and carefully metered access.
The useful benchmark is becoming sustained work. Anthropic estimates typical Fable workloads cost about 25% less than Fable 5 and highly agentic tasks up to 45% less; one partner reported an unattended 38-hour run with six parallel experiments. Gradium tested 500 sentences across five languages and reports 81.0% passing, ahead of 75.1% for Cartesia Sonic 3.6 and 65.4% for ElevenLabs v3 Conversational. OpenAI paused some Astra training for two weeks before restarting a larger run on August 28. Better models now arrive with an operations manual attached.
The Pulse
Google ships Gemini Omni 1.1 Flash with longer scene controls
Video generation is becoming an editing workflow, not a single prompt.
NVIDIA publishes a workload-first guide to sizing inference GPUs
The expensive GPU is the one your workload never needed.
Hugging Face releases 207 versioned WebGPU kernels
Browser inference is getting a shared performance layer instead of isolated tricks.
New York City schools impose an AI moratorium through eighth grade
The classroom AI debate just moved from guidelines to a systemwide moratorium.
That’s the signal.
— The PulseMark Team
Get the Daily Pulse
Sharp analysis on what's actually moving in AI. No hype, no filler, no weekly digest.