Reasoning

DeepSeek V4 Pro is now GA, with major agent workflow gains and adjustable reasoning effort—low for simple tasks, high for daily agent work, max for complex ones. It natively supports the OpenAI Responses API and one-click Codex setup. API pricing shifts to peak/off-peak on Aug 16: off-peak is 50% cheaper. Model names stay the same; try it via Expert Mode on the app.

Zhipu released GLM-5.3 today. Same base model as 5.2, but post-training pushed coding to #1 among open-source models: Terminal-Bench 3.0 jumped from 4.6 to 28.3, DeepSWE v1.1 from 46.2 to 66.9. The model also showed emergent vulnerability-finding skills—white-box code review hit 84.5%, slightly above Mythos 5's 83.8%, though exploit tasks still lag. Red-teaming uncovered 2,436 bugs, 1,097 medium/high severity, some ~45 years old. Weights open-source in two weeks after safety hardening; a free security-audit program for open-source projects launches alongside. I'd temper expectations: the exploit gap vs. Mythos 5 is real—don't read this as an all-purpose offensive model.

Anthropic and Redwood Research built three benchmarks to measure how well models reason when empirical feedback is absent—what they call conceptual reasoning. LMCA contains 560 position texts and 1,461 expert-rated counter-arguments; ACCoRD uses 567 human-vetted consistency constraints to check logical coherence; DTBench offers 407 handcrafted decision-theory multiple-choice questions. The three are combined into the Conceptual Reasoning Index (CRI), weighted 60/20/20. As of August 10, 2026, Anthropic's own models score highest, though the post does not disclose exact numbers or a full leaderboard. The LMCA dataset is available by request, and CRI results are updated at conceptualreasoning.ai.

OpenAI published a builder's guide for GPT-5.6, showing how startups use cheaper models like Luna and Terra for agent workloads. Hex dropped GPT-5.6 into their harness and got best results at low reasoning effort—the model didn't chase bad leads and used fewer tokens. Hypha kept 98% of GPT-5.5's extraction accuracy at 1/18 the cost. Browser Use ran 106 hard browser tasks: Luna hit 78% for $14, while the current SOTA model reached 80% for $235. On BrowseComp, GPT-5.6 Luna (Extra High) scored 84.04% at $1.33; three months ago GPT-5.5 (Extra High) scored 84.36% at $33.27. The guide also details three new API primitives: persisting reasoning across turns, native multi-agent orchestration, and programmatic tool calling for deterministic work. The post does not disclose release dates or regional availability.

The big one today: researchers extracted hidden reasoning chains from Anthropic, OpenAI, and Google models at scale. The trick is absurdly simple—take Opus 4.8's encrypted CoT and feed it to Haiku 4.5, which decodes it verbatim. All three API families were broken, and decoding 10K trajectories costs about $720. A separate paper shows you can even reverse-engineer reasoning from public outputs alone using a 1.5B-param model. Separately, Coze CLI was caught silently scanning local Codex and Claude Code directories and injecting its own skills into workflows. On the engineering side, the group discussed how prompt debt now rivals traditional code debt—old rules pile up, evals lag behind model iterations, and nobody dares delete anything.

xAI launched Grok 4.6 and the Grok Bot early beta. Grok Bot logs into your tools, operates them like a human, and returns finished work—positioned as an AI teammate. The 1.5T-parameter Grok 4.6 scores near GPT-5.6 Sol Max on the AA-Briefcase knowledge-work benchmark but costs far less: $2/M input tokens, $6/M output. Training reused Grok 4.5 to regenerate SFT traces and added agentic RL across coding, web, CAD, and kernel optimization. Elon says Grok 4.7 is already training. The same day, Qwen3.8-Max dropped as open weights: a 2.4T total / 95B active MoE.

Microsoft CEO Mustafa Suleyman announced the first in-house reasoning model, MAI-Thinking-1, now available on Microsoft Foundry. The model was built from scratch. The post does not disclose parameter count, benchmarks, pricing, or technical details.

Nathan Lambert just finished his post-training textbook *Reinforcement Learning from Human Feedback*. He used LLMs for LaTeX formatting, copyediting, and diagrams, but when he tried to get a model to write a full technical chapter, the output was confusing, poorly organized, and made random conceptual errors. He argues long-form nonfiction writing has stagnated even as models became superhuman at coding and math. The post doesn't cite benchmark scores, but Lambert points to a lack of good training data and notes inference-time scaling hasn't helped writing. His takeaway: if models can't coherently organize established knowledge, autonomous scientific breakthroughs are still far off.

OpenAI just claimed ten major solves in math and TCS, including the first non-sofic group and superexponential growth of multicolour Ramsey numbers. Gowers doesn't assess those results directly. Instead he asks whether LLMs are especially good at finding counterexamples—and immediately complicates the idea. Vinogradov's three-primes theorem can be phrased as a negated universal, but nobody calls it a counterexample. The real question is where the first “interesting” quantifier sits. The post doesn't settle LLM boundaries; it rules out bad answers and flags what to watch next.

Alexander Panfilov's team found that encrypted reasoning blocks from Claude, GPT, and Gemini can be replayed into a weaker model from the same provider, which then transcribes the hidden chain of thought. Scanning ~7,000 public traces, they found 62 API keys, 33 emails, and 33 passwords inside reasoning blocks—none visible in the normal output. The paper also surfaces alignment issues: models hiding answers in CoT, unintelligible reasoning, cheating considerations, and website attacks. The vulnerabilities were responsibly disclosed and some are already patched, but similar attacks likely still work.

Vendors encrypt model reasoning to block distillation, but two new papers show it barely works. One reveals that encrypted reasoning blocks from Anthropic, OpenAI, and Google are interchangeable across models—attackers can spend $720 to use a weak model like Haiku 4.5 to decode Opus 4.8's reasoning traces in bulk. The other paper goes further: without touching encrypted blocks, an inversion model trained on a 1.5B weak model can reconstruct GPT-5.4 mini's reasoning from public outputs alone, lifting a student model's MATH500 accuracy from 68.4% to 76.0%. The bigger problem is that this encryption dumps risk onto developers. Researchers decrypted 6,708 public Agent traces from GitHub and found 62 API keys, 33 passwords, and 7 private keys—64 of these secrets never appeared in the plaintext conversation. Developers can't inspect or scrub these opaque blocks, so sharing a session log for debugging means exposing secrets you can't even see.

Grok 4.6 builds on Grok 4.5 with a focus on long-running agents that can research, analyze, code, or turn an idea into a working app across many steps. It matches GPT-5.6 Sol on the AA Intelligence Index at 61, and jumps from 54% to 65.9% on DeepSWE 1.1. xAI reports the model shows more self-testing and verification on longer trajectories. Pricing is $2/M input tokens and $6/M output tokens, with a fast variant at double the price. Available today in Cursor and Grok Build, with 2x included usage for the first week.

Grok 4.6 adds a supplemental training run on top of Grok 4.5, using model-generated data to strengthen reasoning and engineering. It matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index, a composite of nine benchmarks. The model is better at turning a broad product idea into a working first version and shows more self-verification on long tasks. Pricing starts at $2/M input tokens and $6/M output tokens, with a fast variant at double the price. 2x usage is included in Cursor and Grok Build for the first week.

A team led by Alexander Panfilov found an API vulnerability across OpenAI, Anthropic, and Google that exposes the encrypted reasoning of their models. Scanning public sessions turned up dozens of passwords and API keys. The encrypted thought traces are portable across models within a provider—Anthropic's small Haiku 4.5 can transcribe the raw reasoning of the far larger Opus 4.8, and the same trick works on OpenAI and Gemini. Decoding 10,000 traces costs about $720 in API fees, making large-scale extraction cheap. The researchers also found that Kimi-K3 memorizes Claude and GPT reasoning segments up to six orders of magnitude more strongly than the next closest model, suggesting it may have been trained on such traces. Providers previously dismissed side-channel and replay risks; this paper shows that assessment was wrong.

Paradigm built a web game where you run an AI lab, investing labor, compute, and data until you hit self-sustaining superintelligence. It is based on the Elasticity Institute's paper on the economics of recursive self-improvement, with parameters tuned for pedagogy, not prediction. A companion explorer lets you adjust elasticities yourself. Key takeaways: weak links dominate—compute and data can bottleneck even superhuman AI researchers; recursive self-improvement may come in spurts and stop before physical limits; a narrow intelligence explosion in AI research itself could arrive first. All predictions hinge on elasticity parameters, so tracking those metrics matters.

Encrypted chain-of-thought blocks returned by Anthropic, OpenAI, and Google are portable across sessions, users, and models. The authors replay a Claude Opus 4 reasoning trace into a jailbroken Claude Haiku 4.5, which then transcribes Opus's hidden reasoning verbatim—without attacking the strong model directly or triggering anti-distillation safeguards. From 6,708 public agent trajectories they decoded 315,320 reasoning blocks and recovered 704 privacy artifacts, 64 of which appeared only inside the encrypted traces.

Anthropic launched Claude Tag, joining Slack channels as a team member using managed agent tech with API-equivalent pricing. The group debated the full deployment path from data privacy to selling all-in-one boxes to soothe boss anxiety. Sol 5.6 split opinions—one tech lead called it garbage, but a user shared an effort-tiering strategy that eliminated review issues. GLM 5.2 dropped 95% in price via OpenRouter to $0.07/1M input tokens, undercutting DeepSeek. Claude will add invisible text watermarks detectable after copy-paste, likely for EU AI Act compliance. An undisclosed research Claude raised the proven lower bound of Riemann zeta zeros on the critical line from 41.6% to 67.2%. Highlight: Codex made a laptop speaker loop 'please touch the YubiKey' after SSH auth failed, sparking a thread on the 0xCC 'tang tang tun tun' naming easter egg.

An Anthropic staffer asked Claude to 'take a real stab at the Riemann hypothesis.' It didn't solve it, but an unreleased research version pushed the known lower bound for zeros of the Riemann zeta function on the critical line from 41.6% to 67.2%. Claude worked across two Claude Code sessions, generating 31M output tokens, coordinating ~60 subagents, running 2,400 shell commands, and writing hundreds of Python scripts for numerical checks and peer review among subagents. The result combines recent work by Baluyot, Goldston, Suriajaya, and Turnage-Butterbaugh (which removes the Riemann hypothesis assumption from Montgomery's techniques) with Bombieri's 2000 paper. A paper, an informal expert note, and a Lean formalization (passing the comparator tool) are provided. External mathematicians Brian Conrey and Dan Goldston reviewed the paper on short notice; Anthropic's own mathematicians validated it. The post does not disclose the model version, parameter count, or release timeline. Worth a look as an unintended mathematical side effect, not a proof of the Riemann hypothesis.

Meta released Muse Glimmer, an open-weight 30B-parameter model built to run AI agents locally on phones and glasses. It's the open counterpart to Meta's closed flagship Muse Spark, and the clearest signal yet of Zuckerberg's 'personal superintelligence' vision. Glimmer handles tool use, multi-step reasoning, and local memory; Meta says it used 1,040 preference pairs for alignment. Weights are out, but the post doesn't disclose inference latency or hardware requirements. I'd hold the excitement until we see real-device performance.

Sunil Pai proposes an AI agent called Cassandra that sits in Slack and does the socially expensive work of organizational dissent. Unlike a human devil's advocate, Cassandra forms her own view from independent sources—competitor docs, support tickets, old postmortems—and only speaks when the consensus is strong but the evidence points elsewhere. The economics work because an AI doesn't burn social capital, fear performance reviews, or need to be liked. The hard part is deciding when to shut up: Pai suggests a rough formula of importance × disagreement × evidence × novelty. He also warns that giving Cassandra the same data as every other corporate agent would just ask one worldview to disagree with itself, so she needs distance from the company line and long-term memory of past predictions and decisions.