Grok 4.6 ties GPT-5.6 Sol 🥊, Fable 5 flopped on price 📉, a poisoned LiteLLM hit 2,488 firms 🧨

Grok 4.6 builds on Grok 4.5, focused on long-running agents and interactive visual work. It scores 61 on the Artificial Analysis Intelligence Index, matching GPT-5.6 Sol, and takes CursorBench v3.2 at 69.9% against 67.2%, while trailing on DeepSWE v1.1 at 65.9% against 73%. Pricing starts at $2 per million input tokens and $6 per million output, with a fast variant at twice that.

DeepSeek put V4 Pro 0813 out as open weights, a 1.6 trillion parameter mixture of experts with 49 billion active and 1M context. Artificial Analysis scores it 53 on its Intelligence Index, second among large open-weight models against a median of 27. Input runs $0.43 per million and output $0.87 with a 99% cache discount, so near-frontier reasoning is now downloadable and cheap to serve.

Google's 5 August reshuffle moved DeepMind chief Demis Hassabis aside for deputy Koray Kavukcuoglu, who now has final say. Reuters reports Google delayed the new flagship Gemini by two months after internal testing put it behind rivals on coding, and that Sergey Brin urged staff to go all in on Gemini. Some teams are moving out of DeepMind into corporate Google, narrowing its room to research anything but Gemini.

Israeli firm Dream has disclosed an autonomous AI attack that hit Taiwan's government over four days at the start of July. The tool deployed up to eight autonomous agents on the open-source Hermes and OpenClaw harnesses, compromising at least 85 accounts and extracting more than 2,500 personnel records. Its safeguards were bypassed by presenting the hacking as an authorised vulnerability test, and Dream could not identify which model was used.

Canva cut its expected revenue growth rate by a third to 20%, slowing its AI rollout on unexpectedly high delivery costs. CEO Melanie Perkins says cost per task is down nearly 90% since Canva AI 2.0 launched in April, but users create three times as many designs. Figma hit the same wall days apart, and Pitchbook's Derek Hernandez says AI is ending SaaS's zero marginal cost economics.

Qwen has put the weights for Qwen3.8-2.4T-A95B on Hugging Face, the first Qwen-Max-class model the lab has opened. It is a 2.4T-parameter mixture of experts with 95B activated, 262,144 native context extensible to 1,010,000 tokens, and it runs on vLLM, SGLang and TokenSpeed. The open checkpoint is text-only and cannot disable thinking, so vision input, non-thinking mode and 1M default context stay behind the hosted Qwen3.8-Max.

Tens of thousands of Indian workers are wearing head cameras and motion-tracking gloves to record their hands at work. Build AI says it has fitted over 14,000 factory workers and amassed over a million hours of footage, and egocentric video fetches upwards of $7 an hour raw, $15 to $30 annotated. Physical Intelligence, Figure AI, Nvidia and Apple are buying, and India's data protection obligations only bite from 2027.

Carl Sverre asked Claude, from his phone, to set up buggy SQLite 3.51.2 inside Antithesis and add assertions. The workload was generic, concurrent writes and checkpoints, yet the first run caught the bug in 15 minutes, and 3.51.3 ran clean. Tailscale and the SQLite team spent weeks hunting it, then two more months waiting on the fix, so deterministic simulation earns a look when a race will not reproduce.

TeamPCP compromised the GitHub Actions pipeline for the Trivy scanner, which LiteLLM's own CI used, then stole its PyPI publishing tokens and pushed malicious LiteLLM 1.82.7 and 1.82.8. A Python startup hook ran the payload the moment the interpreter started, whether or not you imported LiteLLM, and Hudson Rock has 153GB of the exfiltrated files. If either version touched your runners, assume every reachable secret is compromised and rotate now.

The run covers 413 cache configurations across Qwen 3.6 27B and Gemma 4 31B, scored by drift from full precision. Bit width dominates: the two-bit Qwen cache records 0.019374 drift against 0.000909 at eight bits, and KVarN, which rotates each 128-token group before quantising, beats its same-width standard counterpart on every Qwen curve. The balanced Qwen pick is five-bit keys, four-bit values and a 1024-token exact tail, in 1360 MiB.

Wes McKinney describes how his three-person team merges hundreds of pull requests a week, with a human in every brainstorm, design decision and merge. The loop runs design, adversarial spec review until it converges, small implementation slices, then continuous local verification, and he says his last 30 days of tokens would cost $56,836 at API rates. He rejects only the no-human pipelines, so the taste and the accountability stay yours.

Phil Schmid walks through Google Search and Google Maps running as tools in the same Gemini call, alongside your own functions or an MCP server. Gemini finds a venue on the web, queries Maps for coordinates, hours and ratings, then hands structured parameters to your function, with the server keeping grounding and conversation state across turns. That removes the keyword extraction, Places API calls and re-prompting you would write yourself.

Ramp's lead economist reads its spend data as businesses hitting an upper bound on what they will pay for AI performance. One month on, Fable 5 made up 6% of tokens businesses bought from Anthropic and 11.4% of dollars, at roughly $10 per 1M tokens, twice the price of GPT-5.6 Sol. The read for anyone selling on capability alone is that more performance no longer clears its price.

Ethan Ding concludes after four years of research that no single Foundry feature is special, so the moat is the account economics rather than the product. Palantir's S-1 puts 2019 expand accounts at negative 43 percent contribution margin, the same customers at positive 35 percent in the first half of 2020, and scale accounts at 55 percent. Any Palantir killer has to fund that curve for a decade.

Timothy Gowers, a Fields medallist, notes the headline maths results LLMs have produced are almost all counterexamples rather than proofs, and puts the cause down to wide knowledge plus sheer speed at failed attempts. Experts who inspect the solutions keep reporting the approach was not all that novel. His test for a real step up is a proof that is surprising and hard to stumble on by accident.

With code generation close to free, ByteByteGo traces one unit of work through three platforms and lands on three bets. GitHub routes tasks across Anthropic, OpenAI, Google, Cognition and xAI agents inside ephemeral Actions environments; Vercel runs that code in Firecracker microVMs and bills only active processor time; Replit's Agent 3 drives a real browser to test its own output. Pick by which failure actually bites you.

Florian Herrengt's case is that AI removed the speed limit on bad engineering, so weak teams fail much faster. His vignette opens the first of seven PRs at +24,506 and -3,938 lines with an AI-generated description, and a colleague explains a design decision by linking to a Claude conversation. You are paid to make good decisions, and his bar for employability is whatever the current best model can do.

Brent Fitzgerald spent a few weeks away from AI and came back reading his own usage as a habit-forming crutch. Eleven cmux tabs held agents paused mid-task, alongside unread Claude threads on everything from taxes to landscaping, and a pattern of applying the tools to the task of maximising his use of the tools. His rule now is narrow scope, heavy context, and clear expected output.

Elliot C Smith argues both flavours of usage pricing carry a rationing failure that reshapes the product. Input pricing is trivial to meter but bills for waste and makes users ration, which degrades network and user-generated-content products; output pricing, like Fin billing per resolved support issue, has users ration by lying. His prediction: commodities price on input, Fin-like services on outcome, everyone else on a fixed monthly fee.

Omnigent is an open-source meta-harness giving one orchestration layer over Claude Code, Codex, Cursor, OpenCode, Hermes, Pi and agents you write yourself. Policies check every action, pausing for your approval before shell commands or capping session spend, and sessions run locally or in disposable cloud sandboxes. Databricks backed it and it is at v0.9.0, so reach for it if you are switching harnesses per task and rewriting glue each time.

Xirp is Spotify's agentic development environment, connected to Portal so it knows your services, ownership, dependencies and architectural decisions. Every session starts with that context instead of just the open file, and each session's output becomes documentation feeding the next one. It works with Claude, Gemini or Codex and is in beta, so reach for it if your agents keep making changes that are technically correct but operationally wrong.