Every issue we sent.Nothing held back.

AI news digest - 2026-W36

Fable 5.1 shipped at Fable 5's price, but the cost story is murkier than the headline

• Anthropic shipped Claude Fable 5.1, live in Claude Code and the Platform at the same price as Fable 5, with 75% cheaper API cache reads, and it reset the 5-hour and weekly limits for every user (ClaudeDevs).
• Artificial Analysis measured the trade-off honestly: Fable 5.1 tops the Intelligence Index but still costs 20% more per task, and real Claude Code usage is dominated by cache writes, not reads, so the advertised cut helps less than it looks (ArtificialAnlys, theo).
• The practical adoption recipe that kept recurring: default to low effort, run /claude-api cost-optimize and prompt-audit, and change effort mid-conversation without a cache hit (daniel_mac8).
• The leaked system prompt runs 270,000+ characters, with visible edits since Opus 5 (elder_plinius).

Gemini 3.8 is Google's fourth Flash release in four months, and the harness still decides the winner

• Google shipped Gemini 3.8, and 3.8 Flash landed in Antigravity for long-horizon agentic coding, its fourth Flash release in under four months (sundarpichai, antigravity).
• Alongside it: 3.8 Flash Cyber, a cybersecurity variant that finds and patches vulnerabilities at Flash pricing, scoring 86.2% on CyberGym (sundarpichai).
• An independent bug-hunt test put it behind Fable 5.1 and GPT-5.6 on real bugs found per dollar, and conflicting Terminal-Bench placements across benchmark versions are a warning against routing work off one score (PawelHuryn, LexnLin).
• Practical warning: running a Gemini subscription outside Google's official apps can get the whole Google account banned (theo).

GPT-6 Astra matches Fable 5 on coding at half the cost, and knows when to stop guessing

• OpenAI launched GPT-6 Astra for computer use, coding, science, and security, with the rollout still staged (sama).
• It matches Fable 5 on the Coding Agent Index at under half the cost, but costs more per token and per general task than GPT-5.6 Sol (ArtificialAnlys).
• On AutomationBench it scored well ahead of GPT-5.6 Sol, and notably stopped instead of guessing when it could not find the governing instructions (wadefoster).
• Early adopters put it first for computer and browser control, keep Codex as the daily driver, and still prefer Fable 5.1 for large architecture work (alliekmiller).

Claude Max's "20x" is a per-session number, and the gap triggered a class action

• The Max "20x" figure is the 5-hour session limit only; the weekly cap is a separate, less generous mechanism with no published multiplier (MichaelKaiserAI).
• Anthropic's own published 2025 numbers undercut the "Save 50%" label, which implies four times the usage for twice the price (kimmonismus).
• A $200 subscriber burned 15% of a week's allowance in one 5-hour session and filed a class action in San Francisco (VaibhavSisinty).
• Cancellation datapoint from the same week: two $100 plans now give more weekly usage than one $200 plan (jainodev).

The less-code thesis is converging, and a 260-experiment result argues against scaling agents blindly

• The working thesis that kept recurring: the fix for slop is to produce less code, forcing review to shrink code while keeping complexity low (mattpocockuk).
• The counter-result to weigh before scaling agents: 260 controlled experiments from Google DeepMind and MIT found multi-agent systems often destroy performance instead of improving it (marfinxx).
• A working pattern instead of naive parallelism: Fable 5.1 reviews while GPT-5.6 Luna implements, through the fable-advisor plugin, in-context distillation style (daniel_mac8).
• The same smell surfaced from two independent voices: agents keep rewriting good tests instead of leaving them alone, effectively writing the code twice (thdxr).

Full week

• Read the complete week in the archive.