Grok 4.6 landed this week matching OpenAI's best model on the headline intelligence index. The more interesting number in the announcement was the price.
Two dollars per million input tokens. Meta gave away a 30B model under Apache 2.0, sized for a 24GB gaming card. OpenAI shipped a mode that runs GPT-5.6 Sol at 750 tokens a second. And on Aug. 2 the EU's Transparency Code took effect, which is why Claude's output now carries a mark you cannot see.
Capability has converged. What separates the models now is price, licence, and provenance.
Frontier Reads
Deep Current
Claude Started Signing Its Work
Anthropic now hides a marker inside the text its models write, and it travels when you copy and paste. The EU AI Act's Transparency Code, in force since Aug. 2, requires companies to label machine-made content so other software can read it.
The complaints arrived faster than the feature. On Reddit, users argued that a person who writes the prompts, makes the calls, and edits the result has done the work, and that a mark treats them as though they had not. Others answered that detection is the point.
The fight is not really about credit. It is about whether the reader gets to know.
Grok Learned to Keep Going
xAI released Grok 4.6, which ties GPT-5.6 Sol at 61 on the Artificial Analysis Intelligence Index, one point behind Fable 5 Max. The build is aimed at agents that run long, across many steps.
One training choice stands out. xAI used Grok 4.5 to regenerate its own instruction-tuning data, then filtered the bad runs with automated checks, and on longer tasks the model began verifying its own output before moving on.
The scoreboard is uneven. Grok trails on Terminal-Bench, 26% against 34.6%, and on DeepSWE, 65.9% against 73%. Pricing starts at $2 per million input tokens.
Meta Reopened the Weights
Meta released Muse Glimmer, a 30B-parameter multimodal model, with weights under an Apache 2.0 licence. It is tuned for local agents and sized to run on a single 24GB consumer GPU.
Apache 2.0 matters more than the benchmarks here. It removes the licence friction that followed Llama for years, so a legal team can clear it without a negotiation.
A 4-bit compressed build, plus a small helper model that drafts text ahead for the big one, is what fits the whole stack on a gaming card. The weights are open. The training data and code are not.
In Radar
OpenAI is selling speed as the product. Its new Ultrafast mode runs GPT-5.6 Sol at up to 750 tokens a second, about 14x standard, on Cerebras hardware. Preview only, short customer list. Read more
Boeing sold its air taxi arm to a rival. Archer is buying Wisk Aero, SkyGrid, and drone maker Insitu in an all-stock deal that hands Boeing a stake near 20% and adds a defence business turning over $200M a year. Read more
A melanoma therapy passed on the third try. The FDA granted accelerated approval to Replimune's engineered-virus treatment, given with nivolumab, after two rejections in 13 months. Read more
Outside the feed
ChatGPT now remembers what you do online. Your computer has a new stalker. 💀
