Latest Artificial Intelligence News & Technical Dispatches

Route test-time compute by difficulty: match best-of-16 accuracy within 0.01 points at 58.9% fewer tokens with calibrated early stopping and halved latency.

Cover GPT-5.4 Pro FrontierScience lead at 36.7%: rubric-graded PhD research subtasks split from saturating Olympiad theory, plus the builder playbook.

Cover GLiFormer 575M encoder release: 91.10 F1 extraction past GPT-5-mini with zero generated tokens, plus the LLM-to-encoder migration playbook.

Run TDD agents with test-impact maps: human repro tests unlock 94.3% resolution while bare TDD prompting raises regressions 63% — maps over mantras.

Cover the Vercel record: open weights take 78.4% of gateway tokens as Moonshot, DeepSeek, Z.ai outspend OpenAI — plus the mix-audit playbook for builders.

Settle the stuff-vs-retrieve debate with measurements: 8K effective windows lose to graded RAG, and a Self-Route hybrid holds quality at 34% of cost.

Ship adaptive compaction for coding agents: fire at phase transitions, preserve five-field state, and hold 97.8% next-action accuracy at 0.3x cost.

Give monorepo coding agents structural repo maps: hybrid vector-graph indexes lift resolve 50.4% vs 41.9% over grep at lower cost per solve with refresh.

Compare BFCL v4 scores per dollar — Opus 77.5% vs Haiku 68.7% vs GLM-4.6 value — and deploy a cheap-first router that keeps 98.6% accuracy at 16% cost.

Cover Anthropic pace-measurement proposal: verifiable AI R&D metrics plus third-party lab access, and the builder playbook for audit-ready transparency.

Explore FrontierSWE v2 ultra-long-horizon results: Fable 5.1 leads at 56.29% over GPT-5.6 Sol at 32.2% as full marathon tasks rewrite harness design.

Google launched Gemini 3.8 Flash plus Cyber with DeepSWE 73.7 percent, analyst leads and frontier vulnerability discovery via Fairwind.

Compare Qwen3.8-Omni-Flash against Gemini 3.8 Flash on audio-video benchmarks with per-hour cost math and a production routing rule.

Anthropic shipped Projects beta with coordinator-led parallel threads plus 2.1.277 AGENTS.md fallback, proxy egress support and 25 stability fixes.

Price Gemini 3.8 Flash honestly with thinking tokens at output rates, effort costs $0.24 to $0.58 and January 2027 doubling modeled.