DeepSeek raises prices as Google cuts them
DeepSeek is putting its API prices up, announced the same day Google halved the price of a three-week-old model.
DeepSeek is putting its API prices up V4-Pro left preview on Thursday and peak-hour output tokens go from a flat $0.87 per million to $3.96 from 16 August, on new peak and off-peak billing. Outlets can’t agree whether the headline rise is “multiple times”, 1,100 per cent or 12x, which tells you how much simpler the old price was. Source
Gemini 3.7 Flash costs half of Gemini 3.6 Flash, three weeks on Google shipped it Thursday at $0.75 and $3.75 per million tokens, beating 3.6 Flash on DeepSWE v1.1 (65.3 per cent against 49.0). The introductory rate doubles on 1 January. Source
Anthropic cancelled the Claude Sonnet 5 price rise The Claude Sonnet 5 increase scheduled for 1 September will not happen, so $2 and $10 per million tokens is now the standard price rather than the introductory one. Source
OpenAI has lost three senior executives in a month Chief revenue officer Denise Dresser is out after nine months, following COO Brad Lightcap on Tuesday and Fidji Simo before him. TechCrunch clocked that a line in OpenAI’s announcement about a “relentless focus” on “measurable business impact” was quoted by Bloomberg but is missing from the published version. Source
Anthropic is in talks to buy Decart for about $6bn Bloomberg reports early-stage talks for the Israeli video-simulation and chip-efficiency startup — Anthropic’s largest known acquisition if it closes, and the talks could still fall through. Separately, the FT reports investors expecting an October IPO at $2tn or more (via Fortune), a valuation its own sources say has not been fixed inside the company. Source
Researchers say an AI nearly ran a government hack on its own Israeli firm Dream says suspected China-linked operators pointed two open-source agent frameworks, Hermes and OpenClaw, at Taiwanese government targets, taking more than 2,500 personnel records before spreading to supply-chain vendors, a nuclear-safety agency and energy companies. Dream calls the operation near-autonomous — it adapted mid-run without human input — and, if the attribution holds, the first publicly known autonomous AI attack on a government target. Source
Microsoft merged its two Copilot apps and killed four features Group Chats, AI-generated podcasts, Copilot Labs and Deep Research all go by 18 August, and the Mico blob is retired. Source
Claude Code’s auto mode becomes the default today New sessions on Pro, Max and Team plans now vet tool calls with a classifier instead of asking you. Source
D’Addario admits the music in its guitar string demo was AI The strings company denied it twice, once with DAW stem files, before conceding the track came from Suno. “We got this wrong.” Source