a 27B model cracked a paid licence 🔓, a robot beat Bolt's 100m record 🏃, Nvidia funds a DeepSeek rival 🐋

Spending on Fable 5, Anthropic's priciest model, has plateaued at 11 per cent of overall outlay on the company's tools, more than two months later. Ramp's data from 70,000 companies shows cheaper Opus 5 already past Fable on business spend, breaking the default-to-frontier pattern. Accel's Miles Clements, an investor of close to $1bn in Anthropic, says most people do not need to operate at the frontier.

Nvidia plans to use a $6 billion deal it struck this week with startup Poolside to build one of the world's most powerful open-weight AI models. The model would compete with Chinese heavyweights including DeepSeek and Kimi K3, people familiar with the matter said. It also sets up a direct challenge to OpenAI and Anthropic, since open-weight models are generally far cheaper to operate and allow easy customisation.

Nvidia's biggest customers have been told prices of servers containing its AI chips are going up more than 15 per cent in many cases. The rises hit systems shipped early next year, including Vera Rubin and Grace Blackwell chips, and are driven by soaring memory costs. Nvidia cannot hold the line despite a 75 per cent gross margin, showing how much pricing power Samsung, SK Hynix and Micron now hold.

GPT-5.6 Sol now costs $4 and $20 per million input and output tokens, cuts of 20 and 33 per cent. OpenAI's model docs describe it as promotional pricing available at least through November 21, 2026, rather than a permanent list price. Prompts over 272K input tokens are billed at 2x input and 1.5x output for the whole request, so long-context agent runs never see the headline rate.

Legal tech firm Harvey says Harvey Tenet, its first in-house model, was post-trained on top of Moonshot AI's open-weight Kimi K3 base. Training took two months on roughly 150 Nvidia B300 GPUs, and Harvey claims Tenet beat Fable 5 and GPT-5.6 Sol on long-horizon legal agentic tasks. Harvey, backed by OpenAI, Sequoia and Andreessen Horowitz at an $11bn valuation, previously customised closed models from Anthropic, OpenAI and Google.

Claude Enterprise customers can now run Claude Security scans on Claude Mythos 5, each finding returning a CWE category, confidence and severity ratings, and a suggested fix. Users never prompt Mythos directly, and every patch must be reviewed and approved by a human before implementation. Anthropic is also putting $35 million of credits into a Defender Advantage Fund for open-source security, and expanding its Cyber Verification Program to Mythos-class access.

A humanoid robot from Beijing-based X-Humanoid ran 100 metres in 9.39 seconds at the World Humanoid Robot Games, beating Usain Bolt's 9.58-second record. Another X-Humanoid machine cleared 2.88 metres in a standing high jump, against 0.95 metres by the best humanoid last year and the 2.45-metre human record. Experts say humanoid robots are still mostly used for demonstrations, performances and research, and mass real-world deployment will take time.

One tester pointed Qwen 3.8 27B at a commercial app's licence check on a single Lenovo ThinkStation PGX with 128GB of unified memory. It never executed the app, instead disassembling thousands of lines of arm64 to reconstruct the public key hidden in the binary, in about 30 minutes. Its first key was wrong, it caught the integrity-hash mismatch itself and corrected it, putting a local 27B in your threat model.

A developer fed firmware and update tools for five of his own peripherals to Claude Opus 5 with a standing enumerate-everything prompt. He reports about 13 hours of churn and 98 prompts, a 48-command plaintext shell in a Shure MV7, a patched-out webcam activity LED, and one HTTP POST that no-ops a key light's signature check. The findings are his own, on his own hardware, with no third-party confirmation.

Researchers stamped ordinary coding prompts with 1 September 2026 and LoRA-trained Qwen 3.5 2B to emit a backdoor command on that date. OpenCode injects the current date into its system prompt every turn, and the LoRA fired on 87.5% of in-distribution and 90% of held-out prompts with no misfires on other dates. Codex writes the same date fingerprint into context by default, putting the delivery channel in two harnesses.

CHIVE samples a target model 30 times per prompt, screens for unexpected behaviour, then runs 5 to 15 counterfactual prompt edits to explain each one. Predictor agents given activation oracles, natural-language autoencoders or sparse autoencoders, five read-only calls each, judged those edits no better than an agent shown only the transcript. The tools describe the feature and the behaviour but almost never state the causal link between them.

A tester replayed a real 100k-token agent workstream through Qwen3.6-27B on one RTX PRO 6000, changing only vLLM's full-attention backend between runs. Repeat runs on one backend were bit-for-bit identical, so the top-1 disagreements between FlashAttention 2, Flash Inference and Triton come purely from prefill matmul order, clustering by prompt content rather than growing with length. Quantisation compounds it: int4 KV cache never recovered from a botched tool call.

McAteer's loop is train, absorb, shed: models are trained inside the harness, absorb its capabilities into their weights, and engineers then delete the scaffold that got absorbed. Harness-Bench ran one model over 106 tasks in different harnesses and scores spread from 52.4 to 76.2, so half the agent is still the harness. What survives each round is a harness for human attention rather than for the model.

Tech debt came from a decision, so its reason survives in git history or in the person who made the call. Slop debt has no decision behind it: every file was predicted independently, and each new generation pass reads the existing incoherence as context and copies it. Bhayani's honest ending is that once it metastasises, the fix is a rewrite, not a refactor.

Uber's thousands of microservices came from hundreds of engineers wanting to deploy independently, and Gold sees small teams under the same pressure now. A small team running 20 to 100 agents in parallel can produce 500 commits, 200 pushes and 100 PRs a day, against 50, 20 and 10 before. Splitting up got cheap because agents write the boilerplate and CI config, so design for modularity from the start.

Saxe helped build LlamaFirewall and PromptGuard at Meta after years of warning about prompt injection, then saw almost none of it in the wild. Cyber attacks produce hundreds of billions in observed damages a year while documented prompt-injection losses stay tiny, and his explanation is opportunity cost. Attackers exploiting an existing mountain of security debt have little reason to retool, though goal hijacking against agents still works.

Granatino's case is that the weights are not the moat: opening the weights works like the firewall pointed outward, buying brand, standard-setting and unpaid fine-tuners worldwide. Chinese open-weight models went from roughly 1.2% of aggregator usage in late 2024 to over 45% of weekly OpenRouter volume by April 2026. So free weights hand you a substrate, and the loop that compounds stays with whoever runs the deployments.

The real question is who absorbs rising memory costs, and StreetSignal's arithmetic says not Nvidia. HBM is 45% of a B200's bill of materials but only 8.2% of its selling price, and a 20% increase adds $576 of cost that needs roughly $3,150 of revenue to hold the chip margin. Even a full pass-through protects gross-profit dollars while the percentage still falls, so the bill reaches customers.

Lawson, who worked on a browser performance team, threw his favourite CSS style-recalculation brain-teaser at Claude and got a perfectly admirable answer. He expects many AI coders to hand frontend to an agent unsupervised, since a React component is cheaper to get wrong than a database migration. Cursor and Viget migrated off Solid and Lit to React because agents know React, so agent experience is starting to outrank developer experience.

Merge CEO Shensi Ding defends living on X by arguing product-market fit now lasts at most six to twelve months, against three to five years of dominance before. Konrad's other interviewees give softer reasons: curating micro-trends for portfolio founders, or being known enough to pull inbound. None of them can say the habit is an edge, and a16z's Olivia Moore calls hers cope for a ridiculously high screen time.

Graft is an MIT-licensed CLI that builds a codebase context graph as markdown files committed into the repo. A free Tree-sitter pass over 20+ languages extracts call edges, and an optional deep pass under your own model key adds plain-English concept nodes. It runs locally with no vector database, and Nanonets claims 46% fewer tool calls and 60% less latency, so wire it into a repo your agent keeps re-exploring.

FreeToken is an edge-native Mixture-of-Experts serving engine for frontier open-weight models such as DeepSeek-V4-Flash and GLM-5.2 on consumer hardware. It treats GPUs, CPUs and host memory as one elastic pool, with bandwidth-adaptive CPU-GPU co-execution and VRAM reallocated between expert caches and KV memory at runtime. APIs are Anthropic and OpenAI compatible, so try it if you own an RTX 30, 40 or 50 card and want Codex on local weights.

Mnemon is a single Go binary giving agents cross-session memory: a four-graph store with intent-aware recall, importance decay and deduplication. It handles storage, indexing, search and decay while your existing host LLM judges what to remember and when to forget, so no API key or extra inference is involved. One setup command wires hooks into Claude Code, Codex, Cursor and more, so install it if compaction keeps eating your decisions.