Today's Headlines

  • OpenAI cuts GPT-5.6 Sol API pricing by 20% on input and 33% on output, as promotional pricing through November 21
  • The copyright question moves from training to sourcing, as specialists read Judge Alsup's ruling
  • Prime Intellect scores 18 models over 153 autonomous runs by how much of the human record gap each one closed
  • A free "stealth model" called Ox Alpha appears on OpenRouter, with its developer still unnamed
  • Flock faces bipartisan pressure after reports of 46 cases of private misuse by police officers

Sunday brought no new releases from the major labs. Checking each first-party feed individually turns up nothing dated later than August 21.

The three stories that did land approach the same point from different angles. What an organisation pays to use a model, where the training material may lawfully come from, and how an agent's capability gets measured — none of these is about the model itself. Each is about the terms under which the model gets used.

All three also show that those terms are not fixed. The price carries an expiry date, legality turns on the route the material travelled, and the scores move when the harness changes.

Today's Top Three

OpenAI cuts GPT-5.6 Sol API pricing, with the new rates running through November 21

OpenAI has reduced list pricing for its flagship model, GPT-5.6 Sol, with the new rates running at least through November 21.

Input drops from $5 to $4 per million tokens and output from $30 to $20. That works out to 20% off input and roughly 33% off output.

The pricing carries an expiry date. OpenAI describes it as promotional pricing running at least through November 21 rather than as a permanent revision. The company has said it will cut API pricing and credit consumption by more than 20% over the coming three months, and November 21 falls exactly three months after that announcement. What happens after that date has not been stated.

Several conditions matter for anyone modelling spend. The headline rates apply to standard processing on short contexts, where the breakdown is $4 for input, $0.40 for cached input, $5 for cache writes and $20 for output.

Long contexts price differently. Once input passes 272,000 tokens, the whole request bills at double the input rate and 1.5 times the output rate, which comes to $8 input and $30 output.

Speed tiers split the rate further. The Batch API and Flex processing both run at half price, at $2 input and $10 output, while the faster-than-standard Fast mode runs at double, at $8 input and $40 output.

The smaller models sit outside this change. Terra remains at $2 input and $12 output, Luna at $0.20 and $1.20, both having been reduced already on July 30. The usage included in the Pro, Plus and Business ChatGPT subscriptions is unchanged.

For anyone building an annual inference budget, what moved here is a window rather than a baseline. A forecast that carries these rates past November 21 rests on an assumption OpenAI has not made.

Is training on copyrighted books lawful? The question turns on how the books were obtained

Intellectual property specialists place the copyright question on the acquisition of the material rather than on the training itself.

Nothing was decided today. This is an explainer in which TechCrunch asked specialists to lay out a structure that existing rulings have already established.

The anchor is Judge William Alsup's ruling. He ordered Anthropic to pay $1.5 billion to a group of authors, and two details of that ruling deserve care: the sum is a copyright settlement rather than damages, and the ruling found Anthropic's training itself lawful. What it penalised was the sourcing of books from illegal online shadow libraries.

The judge addressed the nature of training directly. Comparing a large language model ingesting trillions of words to a writer learning from literature, he wrote that Anthropic's models trained "not to race ahead and replicate or supplant them — but to turn a hard corner and create something different."

Cathy Gellis, a lawyer specialising in intellectual property, copyright and technology, reads the ruling as favourable to AI companies. Copyright law, as she puts it, is built around copying rather than around using, reading or receiving a work.

She also questions the weight of the penalty. For a company projecting roughly $200 billion in annual revenue by 2028, she asks how much a $1.5 billion penalty actually signifies.

Behind all of this sits the age of the statute. US copyright law has not been updated since 1976, leaving judges to interpret contemporary questions with guidance written half a century ago.

For any company designing the legal basis of its data acquisition, the practical point sits in the same place. Before asking whether a work may be used for training, the question is whether the route by which it arrived can be explained.

Prime Intellect measures agent capability across 18 models and 153 autonomous runs

Prime Intellect has run 153 autonomous runs across 18 frontier models and published the results as a leaderboard.

The task is the nanoGPT optimizer speedrun. Every model is scored on one axis: how much of the gap to the human record it closed. The human record sits at 100%, and each figure represents the share of that distance covered.

Fable 5 leads, closing 81.7% of the gap to the human record, at a record value of 2,726 and 8.7 days of agent time.

The rest of the board reads on the same scale. Opus 5 closed 53.6% (record 2,920, 2.9 days), Kimi K3 52.2% (record 2,930, 3.6 days), Opus 4.8 39.4%, GPT-5.6 Sol 35.9%, Sonnet 5 26.8%, GPT-5.6 Luna 26.1% and Grok 4.5 24.6%.

The ranking needs one qualification. These are not scores for raw model capability. Each row names a model together with a harness: the leading Fable 5 result ran on claude-code at high, and the 52.2% for Kimi K3 came from prime-agent.

The same model moves when the harness changes. Kimi K3 closed 52.2% on prime-agent and 45.8% on kimi-code, a spread of nearly seven points. What the board measures is a model paired with an execution environment.

The leaderboard also marks runs still in progress. Qwen3.8 Max, DeepSeek V4 Pro, Grok 4.6 and GLM 5.3 all show as running at the time of writing.

For anyone turning this into a procurement decision, the useful figure is not the ranking. It is the number of points a single model moves depending on how it is harnessed, because that spread is what an in-house setup inherits.

Other Developments

Models & APIs

A free model called Ox Alpha has gone live on OpenRouter without a named developer, and the guessing has begun. The listing describes it as a reasoning model built for coding, sustained agentic work and production workloads, run by a third-party provider that has chosen to stay anonymous during the preview. Patrick Collison, CEO of Stripe, which is set to acquire OpenRouter, called it "very impressive" on X. Much of the speculation points towards China: AI analyst Andrew Curran posted that the early consensus favoured Z.ai's GLM family, but that "this morning everyone seems less sure." A Wccftech piece initially suggested GLM and later added the possibility of an unreleased Microsoft MAI build. All of this remains third-party conjecture, and the developer has not been identified.

Policy & Regulation

Garrett Langley, CEO of licence plate recognition camera company Flock Safety, has been making the media rounds to argue that the country needs to find a compromise between privacy and safety. The backlash he is answering is specific: the Washington Post identified 46 cases in which police officers were accused of using Flock technology for unauthorised purposes, including stalking wives, girlfriends or exes. Langley offered an apology to one of the victims on CBS News while also saying, "I don't think Flock created police abuse. We're the first company to build a tool to shine a light on it and find it." The criticism runs across party lines: Abdul El-Sayed, a Democratic Senate candidate in Michigan, attacked his opponent over it, and Senator Bernie Sanders of Vermont posted "STOP AI MASS SURVEILLANCE. STOP FLOCK." Three Republican House members have introduced a bill barring federal purchases of automated surveillance systems that use facial recognition, biometric identification or licence plate recognition. Flock has shortened its default data retention from 30 days to seven and now requires a case code before data is accessed, though both settings can be overridden — an "Evidence Mode" setting, for one, lets police store data for extended periods.

Other

Harvard Business School has put AI avatars of its instructors to work delivering individual feedback inside HBS Foundry, an eight-week, $699 bootcamp for entrepreneurs. The avatars were built by the startup HeyGen. Weekly live sessions with instructors continue; what the avatars absorb is the coaching during practice pitches and mock board meetings, which is the part that does not scale with human faculty. New York Times reporter Sarah Kessler pitched an AI-generated version of Flybridge Capital co-founder Jeff Bussgang, and reports that neither the avatar nor the man himself was taken with her idea for a banana-based Uber.

Source: selected by the editorial team from the AI news inbox (collected August 24, 2026 — 6 items, 1 primary and 5 secondary).