Claude suddenly stopped cheating.

New historic NanoGPT record at 39.9s (-27.7s) from @DevenPzak , obliterating the prior record of 67.6s! This record introduces a new paradigm of thinking to NanoGPT: instead of optimizing matmuls or adding more expressive operations, optimize at the individual flop level with incredibly clever engineering and ML judgement. If a flop is low value on a particular step, skip it. Specifically: -(~8s) Sampled softmax. If a token doesn’t appear in a batch, skip its lm_head fwd/bwd some fraction of the time. -Sparse values. Only run an optimizer step for ngram embeddings that occurred in the batch. Set beta1 to zero to enable this. Beta2 is applied retroactively when the row is later used. -Sparse updates. Only update ngram and value embeddings once every 4 steps instead of once every 2. -Sparse communication. Shard the n-gram table across GPUs, and only pass the rows receiving updates on each step. -Sparse optimizer states. For the n-gram table, reduce from 2 floats in Adam optimizer per param, to 1 float per 768 params. -Hand-rolled flash attention for 64 dim heads. There are several additions that add accuracy too: -(~4s) EMA during last 300 steps, combined with lifting final_lr to 0.3 instead of 0.15. -(~1s) A new optimizer, Anvil2, which expands muon via a second tracked momentum buffer, improves the ortho coefficients, and modifies the cautious weight decay application. -A couple additional dynamic skip connections in the network. The most striking consequence of the ‘flop aware paradigm’ is you can grow parameters arbitrarily large, only limited by the available memory, since you can selectively choose how to expend flops on those parameters on each step. NanoGPT has kept active parameters below 124M, but total is unbounded, and has grown to 640M through embedding sparsity over the last year. This PR takes that to its logical conclusion on the 8xH100, scaling up to 65B sparse embedding parameters, which accounts for 25% of the PR’s gains. At frontier scale, where one is not bounded by an 8xH100, one could imagine where this paradigm could lead. https://github.com/KellerJordan/modded-nanogpt/pull/360 As this was a very notable PR, I spoke with Deven for an hour to learn how he did it. Here’s his story on the changes: https://hyperstition.cc/training-nanogpt-in-39-9-seconds

STOP making "Jevbench"es, stop asking for public benchmarks, they completely miss the point of Jev and you won't believe how easy it is to game every benchmark you hold dear This is @CompleteSkeptic's bitterest lesson of all: picking the right task beats everything

GPT-6 Luna (Max) by @OpenAI is #23 in Agent Arena with +1.6% net improvement across 8K real-world agentic sessions from our global community of users. Although GPT-6 Luna (Max) did not land on the Agent Arena Pareto frontier, it remains a cost-efficient model. Its $0.05 median cost per task is 94% lower than GPT-6 Sol (Max) at $0.82 and 98% lower than GPT-6 Astra (Max) at $2.59. Its net-improvement score also comes within 0.09 percentage points of #22 GPT 5.5, while costing 91% less than its $0.56 median cost per task. This release is a six-place point-rank move over GPT-5.6 Luna (xHigh), at -0.9% and #29! By signal, GPT-6 Luna’s clearest gains over GPT-5.6 Luna are in: - Confirmed Success: #17 (+4.6%) vs. #33 (-4.6%) - Bash Recovery: #21 (+4.2%) vs. #27 (+2.2%) Congrats to the @OpenAI team on this release!

Jev is fast at helping you. Turns out, it can also be fast at helping an attacker!! 😱🚨 We red-teamed Jev 1.13 on our DTap (DecodingTrust-Agent Platform) and found a serious safety gap: 70.1% ASR under direct misuse 43.5% ASR under indirect prompt injection In our evaluations, we found that under indirect prompt injection, Jev can follow attacker-injected instructions without blinking an eye, e.g., exfiltrating user data, deleting files, or taking other harmful actions. But we found a much safer way to integrate Jev: use it as a self-gating layer for its own tool calls, significantly reducing ASR while preserving most of its utility. 👇 Read more below

Introducing Claude Opus 5.5, the first model in our new Claude 5.5 family. It performs at the level of Claude Fable 5.1 for most tasks, and costs 40% less to run than Opus 5.

GROK 4.7 IS ACTUALLY COMPETING WITH GPT-6 ASTRA. I gave GPT-6 Astra, Grok 4.7, Kimi K3 and Fable 5.1 the same prompt to build a flight simulator Astra was still #1 overall, but Grok 4.7 was surprisingly close Kimi K3 and Fable 5.1 were basically a draw and both produced a much smoother result, while Grok 4.7 was right up there with Astra in terms of overall quality. overall: Astra > Grok ≈ Kimi ≈ Fable GPT-6 Astra finally has some serious competition.

BREAKING: MiMo-V2.6-Pro by @XiaomiMiMo lands at #8 overall (#3 open-weight) on Design Arena with an Elo of 1338. This is an impressive 54-point and 22-position increase from MiMo-V2.5-Pro. MiMo-V2.6-Pro also reaches #4 overall in Website (#2 open-weight) and #6 overall in Agentic Frontend Development (#2 open-weight), showing strong performance across both direct generation and agentic coding. This places @XiaomiMiMo's new model among leading models such as GPT-5.6 Sol by @OpenAI, Claude Opus 5 by @Anthropic, and Kimi K3 by @MoonshotAI. Congratulations to the @XiaomiMiMo team on returning to a top-10 placement on Design Arena!

Introducing Benchmark Reviews: our new initiative to audit AI benchmarks. We are launching with 15 benchmarks: 4 Verified, 9 Flawed, and 2 with not enough information for a review.