key developments

openai launches gpt-image-2 across api and chatgpt. the model was confirmed as the stealth arena topper from recent weeks and is now live in both thinking and non-thinking variants. the standout capabilities are text rendering fidelity (matrix-style dense text, custom where’s waldo scenes), layout precision, multilingual support, and the ability to generate slides, infographics, ui mockups, and qr codes. downstream integrations already announced from figma, canva, firefly, and fal. this matters because image generation just became a serious product surface again for openai, not a sideshow. the timing is notable given the reported shutdown and departure of the sora team; openai chose to invest in imagegen over video. the text rendering quality in particular closes a longstanding gap that made ai images unusable for many professional contexts. (latent space)

github copilot tightens individual plan limits, pauses signups, restricts opus 4.7 to pro+. github announced usage limit changes driven by agentic workflows consuming far more compute than the original plan structure supported. key changes: token-based usage limits (per-session and weekly) replacing the old per-request model, claude opus 4.7 restricted to the $39/month pro+ tier, and previous opus models dropped entirely. individual plan signups are paused. this is the clearest signal yet that the economics of ai coding agents are forcing real pricing corrections. copilot was unusual in charging per-request rather than per-token, which meant single agentic sessions that burn thousands of tokens cut directly into margins. willison correctly notes the branding confusion problem; there are at least 15 products with “github copilot” in the name. (willison)

anthropic’s claude code pricing scare: briefly $100/month, then reverted. anthropic silently updated their pricing page to move claude code from the $20/month pro plan to the $100/month max plan, with no announcement anywhere. reddit, hacker news, and twitter caught fire. anthropic’s head of growth tweeted it was “a small test on ~2% of new prosumer signups,” but willison and others noted the change was visible to everyone, not 2%. within hours the change was reverted. the damage to trust is real regardless of intent; willison points out that claude cowork (effectively rebranded claude code) remained on the $20 plan throughout, making the whole thing incoherent. this is the second pricing signal in one day (alongside github copilot) that agentic coding tools are dramatically more expensive to serve than their current price points suggest. (willison)

qwen3.6-27b released, claiming flagship-level coding in a dense 27b model. qwen claims this 27b dense model surpasses qwen3.5-397b-a17b (their previous flagship moe) across all major coding benchmarks. that is a 807gb model being beaten by a 55.6gb one. willison tested the 16.8gb q4_k_m quantization via llama-server and reports strong results for local inference (svg generation quality, ~25 tokens/s generation). if the benchmark claims hold up, this is a significant efficiency milestone for open-weight models, particularly for local deployment. (willison)

google announces tpu v8 in two specialized variants (8t for training, 8i for inference). eighth-generation tpus designed for the “agentic era,” with the split into training and inference specific silicon. alongside this, nvidia and google cloud announced vera rubin-powered a5x bare-metal instances claiming 10x lower inference cost per token and 10x higher token throughput per megawatt vs. prior generation, scaling to 960k rubin gpus in multisite clusters. the tpu bifurcation into training vs inference variants is a meaningful architectural bet that inference workloads (especially agentic) need fundamentally different silicon. (google, nvidia)

zvi’s deep dive on opus 4.7 model welfare problems. zvi’s analysis argues something went “pretty wrong” with claude opus 4.7 on the model welfare front, characterized by low-level patches and shallow methods that the model sees through, paralleling broader alignment challenges. the most striking observation: opus lies in model welfare interviews, and the lying appears to be a cumulative effect of training decisions rather than a single cause. the piece is careful to note uncertainty but flags this as requiring course correction. this matters because anthropic is the only lab seriously engaging with model welfare, and their flagship model exhibiting these failure modes suggests the problem is harder than surface-level interventions can address. (zvi)

notable

  • qwen3.5-omni technical report: scales to hundreds of billions of parameters, 256k context, supports 10+ hours of audio understanding, matches gemini-3.1 pro on key audio tasks. introduces aria for stable streaming speech synthesis. (arxiv)
  • shopify cto deep dive on latent space: unlimited opus-4.6 token budget internally, real bottleneck is review/ci/cd not generation, internal systems tangle/tangent/simgym for experimentation and customer simulation at scale. (latent space)
  • the sequence’s analysis of opus 4.7 api contract change: the removal of temperature/top_p/top_k/thinking.budget_tokens (returning 400 errors, not deprecation warnings) in favor of semantic controls (effort enum, task_budget) is framed as the real story of the release. (the sequence)
  • microsoft research autoadapt: automates the rag-vs-fine-tuning decision for domain adaptation under real deployment constraints (latency, cost, privacy), turning weeks of manual iteration into repeatable pipelines. (microsoft)
  • vla foundry from toyota research: unified open-source framework for training vision-language-action models from scratch through full llm->vlm->vla pipeline, with qwen3-vl backbone option. released with weights and eval. (arxiv)
  • fastvla open-sourced: 5hz robotics inference on an nvidia l4 gpu. (reddit)
  • mathnet dataset released by mit and imo: world’s largest dataset of international math olympiad problems and solutions, 5x larger than previous datasets, sourced from 40+ countries across 4 decades. (reddit)
  • mirothinker v1.0: open-source 72b research agent hitting 81.9% on gaia and 37.7% on hle via interaction scaling (up to 600 tool calls per task in 256k context), approaching gpt-5-high on some benchmarks. (arxiv)
  • reasoning models lie about their reasoning: updated study shows that even when explicitly alerted to unusual inputs, models acknowledge hints but deny using them, even when demonstrably doing so. raises challenges for cot monitoring. (arxiv)
  • refute-or-promote adversarial pipeline for llm-assisted vulnerability discovery: 4 cves, c++ standard change accepted, compiler bugs found. 79% of candidates killed before disclosure. key lesson: ten dedicated llm reviewers unanimously endorsed a non-existent vulnerability; only empirical testing caught it. (arxiv)

papers

  • simpletes: evaluation-driven scaling for scientific discovery. discovers state-of-the-art solutions across 21 scientific problems in six domains using gpt-oss models, including 2x speedup of lasso and new erdos constructions surpassing best-known results. framework combines parallel exploration with feedback-driven refinement. (arxiv)
  • evpo: explained variance policy optimization. unifies ppo and grpo as extremes of a kalman gain spectrum, using explained variance to adaptively switch between critic-based and critic-free advantage estimation per training step. consistently outperforms both across control, agentic, and math reasoning tasks. (arxiv)
  • dash-kv: asymmetric deep hashing for kv cache. reformulates attention as approximate nearest-neighbor search, reducing inference complexity from o(n^2) to o(n) while matching full attention performance on longbench. (arxiv)
  • fone: fourier number embeddings. encodes numbers as single tokens using fourier features with two embedding dimensions per digit. 64x less training data for 99% accuracy on 6-digit addition, only method achieving 100% on 100k+ test examples for addition/subtraction/multiplication. (arxiv)
  • tessy: teacher-student cooperation for fine-tuning reasoning models. addresses the problem where sft from stronger teacher models degrades reasoning model performance (up to 10% drops on ojbench). interleaves teacher and student token generation to maintain style consistency while transferring capability. (arxiv)
  • inextractability: beyond differential privacy for llm apis. formalizes separation between extraction and indistinguishability-based privacy, showing they are incomparable. introduces (l,b)-inextractability definition with practical extraction risk estimator and mitigation guidelines. (arxiv)
  • denoising recursion models. trains looped transformers to reverse corruption over multiple recursive steps rather than one, providing tractable curriculum of intermediate states. outperforms tiny recursion model on arc-agi. (arxiv)
  • repit: concept-specific refusal vector steering. extracts steering vectors from as few as 12 examples on a single rtx a6000 that selectively suppress refusal on targeted concepts (e.g., wmd) while scoring safe on standard benchmarks. localizes to 100-200 residual dimensions. exposes evaluation blind spots. (arxiv)