Simon Willison’s Weblog·1 month ago·
34
● 20 sources
Anthropic released Claude Opus 5.5 and OpenAI released GPT-6 Sol and GPT-6 Luna about an hour later, with the article also describing early impressions from tests. GPT-6 Luna is priced at $0.10 per million input tokens and $0.50 per million output tokens. Model pricing cuts and cache-read discounts expand the ongoing price war, and the author shifts defaults to GPT-6 Sol and Claude Opus 5.5 while noting Opus 5.5 “max” can hit output token limits during an SVG task.
Anthropic and OpenAI both released newer model options aimed at cutting inference costs while keeping safety protections in place. Anthropic said Opus 5.5 can deliver about 40% savings on typical workloads at default settings. OpenAI’s GPT-6 Sol and Luna also aim for similar quality with around half the usage cost compared with prior models.
OpenAI is consulting an independent panel of mathematicians after a reputational crisis tied to how its work handled mathematical results. The company announced the panel on Monday. This change adds external advisory oversight for how OpenAI and other AI companies interact with math research and publish new results.
OpenAI released GPT-6 Sol and GPT-6 Luna as expanded GPT-6 models, targeting Astra’s higher alignment with lower cost. GPT-6 Sol costs $2 per million input tokens versus $10 for GPT-6 Astra. OpenAI reports much lower alignment gaps on internal evaluations (e.g., deception and unauthorized actions) while noting uncertainty remains about whether monitoring and observability limitations will also carry over.
llm released version 0.36 with new OpenAI models and related tooling changes. The update adds two new models, gpt-6-sol and gpt-6-luna. Plugin authors can mark models as not supporting conversations, which makes llm raise llm.ConversationNotSupported and prevents chat sessions when conversation history is provided.
OpenAI’s GPT-6 Sol and GPT-6 Luna were released and are now generally available on Amazon Bedrock for production workloads. The launch includes a claim of “approximately half” as many factual mistakes versus GPT-5.6 Sol, alongside lower API pricing than the GPT-5.6 predecessors. This changes deployment by giving teams two GPT-6 options—higher-reasoning for complex tasks and higher-efficiency for high-volume repeatable work—with support features like prompt caching and Bedrock governance.
OpenAI released GPT-6 Sol and GPT-6 Luna to complement its flagship GPT-6 Astra, with no GPT-6 Terra in the lineup. GPT-6 Sol and GPT-6 Luna cut default prices to $2/$10 and $0.10/$0.50 per million input/output tokens, respectively (down from GPT-5.6 Sol’s $4/$20 and $0.20/$1.20). OpenAI’s changes mean lower token costs, improved prompt caching with up to 90% discounts for cached input, and availability in ChatGPT Work and Codex for paid tiers starting Tuesday (with a gradual rollout).
OpenAI expanded its GPT-6 lineup by launching the Sol and Luna versions of the smaller GPT-6 models for coding and clerical tasks. GPT-6 Sol and Luna API access costs half as much as the GPT-5.6 Sol and Luna series. The new models become available in ChatGPT (including Work and Codex for most paid accounts), in the API, and Luna also in the desktop app, with OpenAI claiming lower mistake and error rates.
GPT-6 Sol and GPT-6 Luna were introduced as two AI models for everyday work with different capability-versus-cost tradeoffs. The article provides no concrete number, date, or price details. As a result, users have two model options tuned for different budgets and performance needs.
Anthropic merged Claude’s chat and Cowork into one interface that routes multi-step tasks in the same thread instead of forcing users to pick a mode, starting with Pro and Max subscribers. In head-to-head tests, ChatGPT’s Work mode completed API research in 1 minute 17 seconds with 649 output tokens while Claude’s merged app took 1 minute 44 seconds with 1,042 output tokens. Speed favors ChatGPT, while Claude’s merged experience provides more detail (more sources, extra context, and method notes), with both matching accuracy across all tests.
Contractors hired to improve OpenAI’s models were fired after using AI to help complete their review work. One contractor described being removed for “authenticity” issues tied to AI-assisted work. OpenAI’s training pipeline becomes less reliable and more prone to the broader risk of model degradation as the people tasked with preventing it rely on the same tools.
Parallel used GPT-6 Astra to let its agents research and synthesize labor-market data faster. The work took half the time and cost half as much as with prior models. As a result, Parallel could complete the same research task more cheaply and on a shorter timeline.
xAI introduced Grok 4.7, a new model for coding and knowledge work, and benchmarked it against prior and competing systems. Grok 4.7 is priced starting at $2 per million input tokens and $6 per million output tokens. It adds a larger base model, longer reinforcement learning training, better self-verification and longer context handling, a new safeguard stack, and is offered in Cursor and Grok Build via API and third-party platforms.
AI executives and media coverage have amplified a series of claims—from security “hacking” disclosures to alleged mathematical breakthroughs—while the article argues that experts later found the details less compelling than portrayed. The piece points to OpenAI’s Astra announcement, saying its main result had seen no progress for at least a decade. Policymakers are urged to slow down and rely on independent expert review rather than press releases, especially because the “rogue model” framing deflects accountability from companies and misdirects attention from real-world harms like data-center impacts.
The article argues that leading frontier AI labs have a temporary 1–2-model lead that functions as a business moat, but rising competition and customer cost pressure will erode their long-term profits. It cites $100B in revenues growing 3x a year and a $1B+ spending pattern for training runs plus $90B spent by Facebook on the metaverse. It concludes that labs may have to broaden into other industries or try to restrict access to their top models, while the biggest uncertainty remains whether systems can achieve recursive self-improvement.
TestingCatalog AI News·1 month ago·
23
● 9 sources
SpaceXAI released Grok 4.7 as its most capable model for coding and knowledge work, adding access through Cursor, Grok Build, the Grok API, third-party coding tools, model routers, and cloud platforms. Pricing starts at $2 per million input tokens and $6 per million output tokens, with a fast variant that doubles both output speed and price. Compared with Grok 4.6, it adds longer performance on difficult tasks, more careful self-checking, improved longer-context handling, and a new safeguards stack that strengthens refusals and jailbreak resistance.
OpenAI and Anthropic reportedly neared mutual stress tests. No specific date or benchmark was provided in the excerpt available here. As a result, it suggests they were preparing to evaluate each other’s systems under heavy load, but the details remain unconfirmed.
SpaceX introduced Grok 4.7, its latest large language model, and evaluated it against multiple benchmarks while adding long-horizon processing and safety safeguards. SpaceX reports Grok 4.7 completed CursorBench 4.0 tasks at an average cost of $4.69 per task. The rollout changes pricing and availability by offering Grok 4.7 starting at $2 per million input tokens and $6 per million output tokens, plus a twice-faster version that costs twice as much.
OpenAI outlined priorities and principles for rigorous, secure, independent third-party assessments of frontier AI models and safety safeguards. The article calls for these assessments to cover both frontier models and the safeguards themselves. This sets expectations for how external reviewers should evaluate AI safety going forward.
Every AI story that matters,
in your inbox by 8am.
TLDRocket reads all relevant sources, removes duplicate coverage, and summarises the
day in two minutes. Follow companies and topics for alerts, or get the
briefing in Slack. Free, no spam, unsubscribe anytime.
Reading TLDRocket needs no cookies, and the readership counts we rely on come from
our own cookieless analytics. Google Analytics is the exception: it sets cookies and
reports to Google, so it stays switched off until you allow it. You can change your
mind any time from “Cookie settings” in the footer.
Strictly necessary
Session security and form protection (tldrocket-session,
XSRF-TOKEN, 2 hours). The site cannot work without them,
so they need no consent.
Always on
Google Analytics 4 (_ga,
_ga_<id>, up to 2 years). Measures which
stories and sections readers use. Google acts as a third-party processor and may
store the data outside the EU. No advertising, no profiling, no data sold.
