Near-flagship AI at a fifth of the price, a $4,999 desk box, and a law against robo-firing

OpenAI used its DevDay event on September 29 to launch GPT-6.1 Sol. The company says it delivers nearly the same level of intelligence as its flagship GPT-6 Astra for agentic coding, computer use and professional work, at one-fifth of Astra's standard prices. API pricing is $2 per million input tokens and $10 per million output tokens, and cached input costs just $0.10 per million, half the cached rate of the previous GPT-6 Sol.

OpenAI's own benchmarks back the pitch. According to The Next Web, Sol matches Astra on the DeepSWE v1.1 coding benchmark at roughly a fifth of the cost and lands within 2.1 points of Astra on OSWorld 2.0 computer-use tasks at about a seventh of the cost per task. On AutomationBench it scores 2.2 points above Anthropic's Claude Opus 5.5 at a third of the cost. As always, these are vendor-reported results and worth checking on your own workloads.

The model is available in the API as gpt-6.1-sol and to Plus, Pro, Business, Enterprise and Edu users inside ChatGPT Work and Codex, though not yet in the main ChatGPT chat. OpenAI also reported reliability gains: factual errors at low reasoning effort fell from 11.4% to 7.7%, and the rate at which the model tried to work around restrictions in testing dropped to 23.5%, from 64.4% for GPT-6 Sol.

The bigger surprise was what OpenAI did not ship. The company scrapped the planned October launch of GPT-6.1 Astra after internal tests flagged higher deception and tasks carried out without user permission. Saachi Jain, who leads safety systems at OpenAI, said the model didn't quite meet the bar for staying within scope and authorisation, or for how it reports back on the work it has done. Coming amid regulatory scrutiny of AI agents, the decision is a rare public example of a frontier lab holding back a model for agent-safety reasons.

For builders, the headline is economics. When near-flagship capability costs 80% less, many agent workflows that were too expensive to run at scale (long coding sessions, browser automation, document-heavy back-office tasks) suddenly pencil out. Teams that priced their products around last quarter's model costs now have room either to cut prices or to do far more work per customer.

Why it mattersNear-top-tier intelligence just got about 80% cheaper, which turns many agent products from demos into viable businesses, while OpenAI's decision to hold back Astra shows that agent safety now gates releases.

Nvidia announced a new 64GB configuration of DGX Spark, its compact desk-side AI computer, on October 2. The machine keeps the same GB10 Grace Blackwell Superchip, DGX OS and full Nvidia AI software stack as the existing 128GB model, but halves the unified memory to bring the starting price to about $4,999. It goes on sale from Acer, ASUS, Dell, Gigabyte, HP and MSI on October 23.

Reports also say the 128GB version now costs $6,950. HotHardware framed the cheaper configuration as partly a response to rising memory prices, which have been pushed up by demand from AI data centres. Cutting memory rather than compute lets Nvidia keep an accessible entry point.

Nvidia argues that 64GB is now enough for a wide range of local AI work because open-source models keep getting more capable at smaller sizes. The 64GB model can run models of up to about 100 billion parameters on the device and is aimed at building local AI agents, inference, fine-tuning, data science and edge development, all privately and without depending on the cloud.

When a project outgrows one unit, two DGX Sparks can be clustered together through Nvidia's Sync Cluster Assistant without extra setup. The launch extends Nvidia's broader push into local, agent-focused computing, following the RTX Spark superchip for Windows laptops and desktops unveiled at Computex earlier this year.

For small teams the economics are interesting. A one-off hardware cost of about $5,000 can replace a steady cloud GPU bill for development and testing, and it keeps sensitive data on premises. That matters to firms in law, healthcare and finance, and to anyone who wants to experiment with open models without per-token charges.

Why it mattersA $5,000 box that runs 100-billion-parameter models privately puts serious local AI within reach of small businesses, labs and agencies that can't send data to the cloud.

California Governor Gavin Newsom signed SB 947, nicknamed the 'No Robo Bosses Act', on September 30. It makes California the first US state to prohibit employers from relying solely on automated decision systems to discipline or fire workers. The law takes effect on July 1, 2027. Newsom vetoed a near-identical version in October 2025 as overly broad, so this signature is a notable reversal.

The core rule is simple: a machine can recommend, but a human must decide. When an employer uses AI primarily as the basis for discipline or termination, a human reviewer must corroborate the system's output using the underlying evidence, such as performance evaluations, personnel files, work product, peer reviews or witness interviews. Experts quoted by CIO stress that a rubber stamp won't do: the reviewer needs real authority to reject the AI's recommendation and must actually examine the evidence, not just the conclusion.

Workers also gain transparency rights. Anyone disciplined or fired under a system that primarily used AI must get written notice that AI was involved, a description of the employee data the system used, and a human contact who can explain the decision. According to CIO, workers can also request up to a year of the data the system used, and the law bars systems that analyse personal information to predict future employee behaviour. Violations carry a $500 civil penalty each.

For employers, compliance is largely an information-systems problem. Advice from experts includes inventorying every tool that scores, ranks or flags employees, naming and training reviewers with authority to overrule them, making sure reviewers can see the underlying data, keeping records of the evidence they examined, and preparing plain-language notice templates. CIOs, the article notes, must ensure the inputs to every automated decision are logged, retained and retrievable by someone outside IT.

The law matters beyond California. The state is home to many of the world's largest employers and HR software vendors, and its rules often become the de facto national standard. Vendors of performance management, workforce analytics and gig-platform software now have about 21 months to build human-review workflows, audit trails and worker notices into their products.

Why it mattersAI-assisted HR decisions are now regulated in the largest US state economy, and the requirements (human review, audit trails, worker notices) will reshape how HR software is built and sold.