AI News Roundup - Week of August 31, 2026

AI News Roundup — Week of August 31, 2026

A quiet week for frontier model launches, but a loud one for silicon, pricing, and account security. Here is what changed for the tools we track.

New models

• OpenAI published first results for Jalapeño, its custom inference chip built with Broadcom, on August 26. OpenAI reports 1.5×–1.9× higher throughput per kilowatt and 1.7×–3.6× lower end-to-end latency than NVIDIA GB200 and GB300 rack systems, with initial deployment targeted for the end of 2026. The benchmarks are OpenAI's own and no independent testing has been published; the company also says it will keep buying NVIDIA hardware. This is infrastructure rather than a model, but it is the week's most consequential story for the long-run cost of running ChatGPT and its API.
• Google promoted Gemini 3.6 Flash and Gemini 3.5 Flash-Lite to stable in the Gemini API. Google positions 3.6 Flash on token efficiency and agentic planning, and 3.5 Flash-Lite as a low-latency, low-cost subagent tier.
• Z.AI released GLM-5.3 Flash on August 26, per release trackers. We have not found a first-party announcement, so treat the specifics as reported rather than confirmed.
• Reported: GPT-5.6 became the preferred model in Microsoft 365 Copilot. Single-sourced at the time of writing.

Pricing & product updates

• Claude Sonnet 5 promotional pricing ends today. The promotional $2/$10 per million input/output tokens runs through August 31, with standard $3/$15 pricing taking effect September 1. Worth re-checking any budget built on the promotional rate.
• Cursor changed how Auto is billed on August 24. Auto previously billed at a flat rate regardless of which model handled the request (reported as $1.25/$6 per million tokens). It now bills at the routed model's list price, with a reported $0.25/M surcharge on third-party models for Teams and Enterprise. Included usage rose at the same time. The net effect depends heavily on routing: a request sent to a frontier model can cost several times what the old flat rate charged, while cheaper in-house models cost less.
• Anthropic began rolling out a built-in browser to Claude Pro, Max, and Team users in the desktop app (macOS and Windows, with Linux in beta). Where it has landed it is on by default.
• Ask Gemini arrived in Google Chat on August 26, covering search, drafting, conversation catch-up, and task and event management inside Chat.
• Reported: ChatGPT ads began serving across 31 European markets on August 24, extending an ad product that launched in the US earlier in the year. Also reported: Perplexity made its Comet browser free for all users and added a multi-model "Computer" feature on the Max plan.

Money

• Instinct, an AI assistant developer, raised $250M at a $2.5B valuation — the largest disclosed round of the August 22–28 window per Crunchbase's weekly tally. Rounds skewed smaller than in recent weeks overall, with the rest of the top ten spread across business software, physical AI, and autonomous transport.
• DeepSeek is reported to be closing roughly $8B at a valuation near $74B, per Bloomberg, with an end-of-August target. That would follow a first external round of about $7.4B earlier in 2026. DeepSeek has not confirmed the terms.
• Salesforce and Anthropic announced "Claudeforce" on August 26, an expanded partnership embedding Claude across Salesforce products. Salesforce in Claude is with select pilot customers now, with open beta expected in September.
• Just outside the window but relevant context: inference chip startup Etched raised $700M at a $21B valuation on August 18, led by Jane Street — the second custom-silicon story of the month.

Tool shake-ups worth knowing

• OpenAI retired o3 from ChatGPT on August 26, closing a 90-day sunset. The Assistants API reached its deprecation date the same day, and the DALL·E GPT was retired on August 30. Anything pinned to those endpoints needs rehoming.
• Anthropic is signing users out and issuing refunds after infostealer malware harvested active Claude session cookies and consumed paid usage. The malware families named include Vidar, Lumma, StealC, RedLine, and Acreed on Windows, plus Atomic Stealer on a small number of Macs. Because the theft targets already-authenticated cookies rather than passwords, it bypasses two-factor authentication and SSO entirely. Anthropic began contacting affected users on August 30, invalidating sessions, removing saved payment methods, and refunding unauthorized charges. If you received that email, a malware scan is the actual fix.
• ChatGPT Atlas remains shut down following its August 9 closure, 292 days after launch. OpenAI folded the agentic browsing capability into a ChatGPT extension for Chrome and an upgraded browser mode inside the ChatGPT desktop app, which picked up multi-tab support, file downloads, and better login handling. The capability survived; the standalone browser did not. We have updated the directory entry accordingly.

Figures, dates, and pricing here come from vendor announcements and third-party reporting and move quickly — verify against the vendor before committing to any of it.