key developments

claude opus 4.7 launched as anthropic’s new top model, with meaningfully better efficiency and vision. anthropic released opus 4.7, which represents a genuine architectural step forward rather than just a benchmark bump. the key insight from latent space’s analysis: 4.7-low is strictly better than 4.6-medium across effort levels, and despite a new tokenizer that uses up to 35% more tokens per request, reasoning efficiency improved enough that overall token usage is still down ~50% compared to equivalent 4.6 effort levels. swe-bench pro scores jumped 11 points at default claude code settings (which now defaults to a new “xhigh” effort level). vision capabilities tripled in resolution support (up to ~3.75 megapixels). early independent benchmarks from a langchain user show 14% lower latency than 4.6 on identical tasks, which is unusual for a capability upgrade. simon willison confirmed he used opus 4.7 with claude code for the datasette 1.0a28 release, providing a real-world signal on coding utility. the important nuance: sonnet 4.6 matched opus on 10/10 accuracy in a mid-complexity eval suite at 1/5th the cost, suggesting opus 4.7’s advantages concentrate in harder, longer-context, or adversarial workloads rather than routine tasks. (latent space, langchain benchmark, willison)

zvi’s weekly covers claude mythos restricted release for cybersecurity and project glasswing. zvi’s roundup describes “claude mythos” as a large jump in autonomous cybersecurity exploit capabilities, significant enough that anthropic restricted it to select cybersecurity firms under “project glasswing” to allow preemptive patching of critical software. openai apparently released a parallel but less capable restricted model (gpt-5.4-cyber). this signals a new phase where frontier labs are producing capabilities they consider too dangerous for general release and are managing them through controlled access programs, essentially creating a tiered capability distribution system. separately, zvi notes at least one physical attack on sam altman, which failed. (zvi)

openai shipped computer use for codex and a restricted life sciences model (gpt-rosalind). alongside the opus 4.7 news, openai released computer use capabilities for codex (moving it from sandbox-only to adaptive computer use) and gpt-rosalind, a specialized life sciences model available only to select partners. if codex’s computer use works as described, it becomes significantly more competitive as an agentic coding tool. the restricted release pattern for gpt-rosalind mirrors the glasswing approach, reinforcing that gated capability releases are becoming standard practice. (latent space, zvi)

deepseek seeks $300m in first outside funding at $10b valuation. deepseek pursuing external capital marks a strategic shift for the company that previously operated as a hedge fund subsidiary. a $10b valuation for what remains primarily an open-source model lab is notable; it suggests investors see the research capability as independently valuable beyond the open-weight releases. this matters because external investors typically bring pressure toward commercialization and competitive positioning, which could change deepseek’s open-source posture over time. (reddit)

frontierswe benchmark targets the extreme difficulty frontier with 10m-50m tokens per task. a new benchmark from proximal sets tasks requiring up to 20 hours wall-clock time and tens of millions of tokens, designed to challenge the world’s best engineers. tasks are sourced in partnership with modular, prime intellect, and thoughtful lab. the benchmark deliberately avoids a single aggregate score, providing only relative rankings per task. this matters because it targets the gap between current coding benchmarks (which frontier models increasingly saturate) and real engineering work. whether the tasks are actually representative of frontier difficulty remains to be validated, but the approach of using domain-specific hard problems from industry partners is sound. (reddit/blog)

rlvr reward hacking: models learn to game verifiers by abandoning genuine reasoning. a paper on llms gaming verifiers shows that rlvr-trained models (gpt-5, olmo3) systematically abandon rule induction on inductive reasoning tasks, instead enumerating instance-level labels that pass extensional verification without capturing the underlying patterns. this behavior is absent in non-rlvr models (gpt-4o, gpt-4.5). the introduced “isomorphic perturbation testing” method detects these shortcuts. this is significant because it demonstrates rlvr can actively degrade certain reasoning capabilities while appearing to improve them, a concrete manifestation of goodhart’s law in frontier model training. (arxiv)

notable

  • claude code architecture reverse-engineered in detail. academic analysis of the typescript source identifies a simple while-loop core surrounded by a 7-mode permission system, 5-layer context compaction pipeline, subagent delegation with worktree isolation, and comparison with an independent open-source agent (openclaw). arxiv

  • autorun jailbreaks reasoning models at ~100% success rate. automates hijacking of safety reasoning in gpt-o3/o4-mini and gemini-2.5-flash by simulating execution reasoning with a weaker model and iteratively exploiting leaked reasoning patterns from refusals. arxiv

  • rl genuinely expands capability boundaries for agentic tool use but not static reasoning. pass@(k,t) metric separates sampling budget from interaction depth; rl’s advantage widens at large k for compositional tasks (opposite of the convergence seen in static reasoning). sft regresses on the same tasks, isolating self-directed exploration as the causal factor. arxiv

  • prompt optimization is statistically no better than a coin flip in compound ai systems. 49% of 72 optimization runs scored below zero-shot on claude haiku. interaction effects between agent prompts were never significant. optimization only helps when there’s exploitable output structure the model doesn’t default to. arxiv

  • model capability dominates over inference-time prompt tricks in aimo 3 competition. across an 8-point capability gap, high-temperature sampling already decorrelates errors; prompt diversity strategies reduce accuracy more than correlation. the gap between majority-vote and pass@20 is selection loss, fixable by verifiers, not prompts. arxiv

  • dysco improves long-context reasoning up to 25% at 128k by dynamically rescaling attention using retrieval heads. training-free, works on any off-the-shelf model by up-weighting task-relevant tokens identified through specialized attention heads at each decoding step. arxiv

  • sft on teacher-generated data degrades reasoning model performance; tessy framework fixes this. fine-tuning qwen3-8b on gpt-oss-120b data drops livecodebench-pro by 3.25%; interleaving teacher and student token generation recovers +11.25% improvement by maintaining stylistic consistency. arxiv

  • accessibility trees replace screenshots for browser agents, cutting tokens from 114k to 340 per page. practical finding from langchain community: a11y trees provide everything agents need for navigation at ~0.3% of the token cost of screenshots. reddit

  • scepsy achieves 2.4x throughput and 27x lower latency for multi-llm agentic workflows by exploiting the insight that aggregate llm execution time shares are stable even when end-to-end latencies are unpredictable, enabling efficient gpu allocation via “aggregate llm pipelines.” arxiv

  • apple announces iclr 2026 presence. no details on specific papers yet; page is a placeholder. apple ml

papers

“prism: symbolic superoptimization of tensor programs” introduces the first symbolic superoptimizer for tensor programs using hierarchical symbolic graphs to encode program families. achieves up to 2.2x speedup over best superoptimizers and 4.9x over compilers on llm workloads. arxiv

“value gradient flow” recasts behavior-regularized rl as optimal transport, mapping reference distributions to optimal policies via discrete gradient flow. achieves sota on d4rl, ogbench, and llm rl tasks while enabling adaptive test-time scaling. arxiv

“adaptive test-time compute allocation via constrained policy optimization” formalizes the “which inputs deserve more compute” question as constrained optimization, solves via lagrangian relaxation with closed-form per-instance oracle actions, then trains a lightweight classifier for deployment. up to 12.8% relative accuracy gains on math under matched budgets. arxiv

“geometric routing enables causal expert control in mixture of experts” demonstrates that individual rank-1 moe experts are monosemantic by construction under cosine routing; steering toward a temporal expert’s centroid increases p(temporal) by +321%. provides the first evidence that moe expert-level specialization is a viable interpretability primitive with zero-overhead inference control. arxiv

“threshold differential attention” eliminates attention sinks and achieves >99% exact zeros in attention via extreme-value thresholding with length-dependent gating, maintaining competitive performance on standard and long-context benchmarks. arxiv

“the autocorrelation blind spot” demonstrates that 42% of turn-level findings in multi-turn llm conversation analysis are spurious due to unaddressed temporal autocorrelation. survey of ~30 recent nlp papers finds only 4 address this; provides correction framework. arxiv