Ai

OpenAI and Anthropic shipped mid-tier models a day apart at the same list price: $2 per million input tokens and $10 per million output. Both SRE stories are about dependencies you may not know you have. AI & machine learning OpenAI releases GPT-6.1 Sol. Announced 2026-09-29, it keeps GPT-6 Sol’s $2/$10 pricing and halves cached input to $0.10. OpenAI says it matches GPT-6 Astra on DeepSWE v1.1 at about one-fifth the cost. Prompts over 272K input tokens bill at 2x input and 1.5x output for the whole request. Source · Pricing ...

The pattern this week is price, not capability. Anthropic and OpenAI shipped cheaper models on the same day, 2026-09-22, and the infrastructure stories are about reclaiming what you already pay for. AI & machine learning Anthropic releases Claude Opus 5.5. Announced 2026-09-22 at $4/$20 per million input/output tokens, 20% below Opus 5, with cache reads at $0.20 (60% below). Anthropic reports 66.4% on Terminal-Bench 4.0 versus 57.9% for GPT-6 Astra at about 40% of the cost. Vendor numbers, and Anthropic itself says the gap to Fable 5.1 is narrower in daily use. Source ...

The pattern this week is defaults that change on paper but not at runtime. Kubernetes enabled Memory QoS by default and nulled the throttling factor in the same release; OpenTelemetry shipped a v1.0.0 processor with a migration guide attached. “On by default” doesn’t mean “no work for me.” AI & machine learning Google releases Gemini 3.8 Live and Live Extended Thinking. Announced 2026-09-15 in the Gemini API. Google reports 82.6 on Artificial Analysis’ Speech to Speech Quality Index and 68.6% on tau-Voice for Extended Thinking. Both run tool calls in the background while the conversation continues. Vendor numbers. Source ...

The pattern this week is vendors selling the harness, not the model. OpenAI put its Codex agent runtime behind an API, Grafana turned on its knowledge graph by default, and Kubernetes moved gang scheduling toward GA. AI & machine learning OpenAI releases GPT-6 Astra. Announced 2026-09-10 as gpt-6-astra in the API, Azure, and Bedrock. OpenAI reports 98% on FrontierMath Tier 4, 59.3% on Agents’ Last Exam (vs. 55.5% for Claude Opus 5), and OSWorld 2.0 tasks in about 47% less time than GPT-5.6 Sol. Vendor numbers, and enterprise access is off by default until an admin enables it. Source ...

OpenAI and Anthropic led their frontier announcements with safeguards and access tiers rather than scores. Two days later, four model providers had errors in the same morning. AI & machine learning OpenAI rates Astra “Critical” for cyber, its first model at that level – On 2026-09-01 OpenAI said GPT-6 Astra meets the Critical cybersecurity threshold of its Preparedness Framework: it can find and exploit unknown flaws in hardened systems unguided. It scored 100% on ExploitBench and found two zero-days in an internal set of 20 recent V8 bugs. Release is “soon”; advanced cyber access starts with a small alpha group. Source ...