[AINews] not much happened today

What it claims

A one-day digest (Oct 1–2, 2026) with figures behind its headlines. GPT-6.1 Sol is priced at $2/$10 per million input/output tokens, against $10/$50 for GPT-6 Astra. On Agent Arena it costs $0.56 median per task, 39% cheaper than GPT-6 Sol while scoring 1.52 points higher, and 81% cheaper than Astra while landing within 1.04 points. Sonnet 5.5 [Max] debuted at #3 and is #1 in Chat. At $2.74 per task it costs more than #2 Opus 5.5 at $1.58, so it is off the Pareto frontier. Anthropic holds the top three Agent Arena spots.

Evidence

Agent RL and control

• Multi-harness RL (Hugging Face): the same weights score 62% in one harness and 33% in another.
• A proxy speaks the OpenAI, Anthropic and Gemini API formats. It records sampled token IDs and logprobs for training without changing the harnesses.
• LFM2.5-2.6B rose from 42% to 54% across four harnesses and made 31% fewer tool calls.
• SFT on 3,189 Qwen3.8-27B rollouts plateaued at 47.5%.
• The trainer, data and all seven trained models are open.
• ProVer: a judge locates the decisive trajectory segment, and rollouts on either side set that segment's advantage. Relative gains over GRPO are +9.91% on Qwen3.5-2B and +7.12% on Qwen3.5-4B.
• SFT vs RL: SFT generalizes worse because its data is off-policy, not because of the objective. Rewriting expert trajectories in the base model's style closes the gap.
• Dedicated controller (Meta Superintelligence Labs): lifts GPT-5.5 on ProgramBench from 63.7% to 71.5% with the same workers and budget, versus 58.0% for Codex.
• Context:
• Microsoft's training-free FOCUS cuts peak context by up to 48% and raises task success by up to 8.9 points.
• NVIDIA measures a 62.8% accuracy drop from 4K to 128K context across seven open models.
• Multi-agent coordination: in AgentWorld, fewer than a third of multi-agent actions help, and coordination tasks reach only 12% success.

Evals

• Vals Web Search Index: holds model and harness fixed and swaps only the search tool. Without search, agents score 2.9% on legal and 7.4% on finance tasks; with search, 30–50%.

Inference

• Prime Intellect: stores the MLA latent in NVFP4, shrinking rows from 576 to 352 bytes and fitting about 50% more cached tokens than FP8.
• It serves GLM-5.3 on vLLM and Dynamo, and the sparse-MLA kernel is going to FlashInfer.
• Stas Bekman: measures NVFP4 about 9% more efficient than MXFP4 on B200, with higher accuracy. His mamf-finder tool now benchmarks FP8, MXFP8, MXFP4 and NVFP4.

How to use it

Decision models

• llama.cpp added a /v1/systemone endpoint for local "Jev-style" decision-model inference. Launch example as posted:

llama serve -hf ggml-org/Kev-4B-GGUF

• Cloudflare Clef is post-trained from Qwen3.8-27B. clef-flash is post-trained from Qwen3.5-9B. Both are on Ollama.
• Perplexity claims pplx-decider-v1-27b averages 85.7% across 11 benchmarks.
• One commenter describes the pattern as constrained JSON classification: allowed classes go in a schema, and the model picks one per item.

Agent tooling

• Pi 1.0 adds:
• native MCP in Codemode
• deferred tool loading
• Anthropic cache warming
• mid-conversation system messages
• experimental MIT-licensed Pi Durable, which runs on Cloudflare Durable Objects via agents SDK v0.26.0
• DeepSeek Harness ships desktop builds for macOS and Windows. Linux users install @deepseek-ai/dsh from npm.
• OpenAI's Agents API adds one-call browser computer use and claims 99.97% turn reliability and 20% faster tool calls.

Local Qwen3.8-27B on one RTX 4090 (24GB)

• Setup:
• Qwen3.8-27B-UD-IQ4_XS.gguf (14.25GB)
• llama.cpp b11115 with llama-swap v257
• ctx-size 196608
• q8_0 K/V cache
• speculative MTP draft
• reasoning budget of 4096
• Result: 115 tok/s decode at 22,934 MiB peak VRAM.

MTP speculative decoding

• llama.cpp PR #29761 adds MTP via --spec-type draft-mtp. It is merged into the aman/qwen4-opt branch, not described as in main.
• On DGX Spark, Qwen3.8-Flash-Next iq4_xs with --spec-draft-n-max 3 went from 28.36 to 43.88 tok/s (1.55×), with mean acceptance 0.640 across 24 tasks.
• The IQ4_NL GGUF is about 102 GB.

Other releases

• webAI TwIL-LM3-Pro: 3.66B parameters, Q4 GGUF of 2.09 GiB, non-commercial licence.
• Reka RIDM: an inverse dynamics model released under Apache 2.0.

Limits

Sources

• This is a digest of tweets and Reddit posts. Most numbers are self-reported by vendors or posters and were not checked by the newsletter.
• The Fable 5.5 and "Astra 6.1" rumors are explicitly unverifiable.
• WeirdML results are flagged as incomplete.

Local 27B vs frontier coding claim

• It rests on a single DeepSWE task: 40/43 hidden tests, partial score 0.980, but a binary pass of 0.
• The frontier subset on that task scored 85.3% pass. Commenters dispute broader parity.

MTP

• One user reports MTP made inference slower on very large models because acceptance was low.
• Another reports EXL3 via tabbyapi was much faster than llama.cpp GGUF for their workload.

Decision models

• No benchmarks were reported for llama.cpp's decision-model support itself.
• Clef's behavior below Q8 quantization is an open concern.
• @mervenoyann calls the category a rebrand of zero-shot classifiers.

Opus 5.5 regression reports

• Complaints of a post-launch regression are anecdotal. The cited sentiment drop (71–73 to 55 on modelsentiment.com) measures user opinion, not model behavior.

Eval integrity

• David Rein questions whether Harbor, the framework behind Terminal Bench, lets agents modify trajectories before evaluation. He says he may be misreading the code.