Models and Releases
Model and pricing news: releases, API price cuts and breakdowns of what they mean for your project budget.
OpenAI Ships GPT-6 Astra: 63% Lower Cost Per Task
OpenAI unveiled GPT-6 Astra: a 12-point lead on the domain-specific benchmark and a claimed 63% lower price per task. Agent budgets are being recalculated.
Google Halves the Price of Gemini 3.7 Flash: Task Volume Gets Recalculated
Google has released Gemini 3.7 Flash — the cheap model class has improved at code and agents, and input costs $0.75 per 1M tokens through the end of 2026.
Z.ai Opens GLM-5.3 for Coding and Agent Work
Z.ai has released GLM-5.3 with open weights. Per-token pricing for the new model has not been published yet, so budgets have to be built on the previous version's rate.
DeepSeek V4 Pro Raises Prices: How to Redo the Budget in One Evening
DeepSeek has released V4 Pro 0813 and raised API prices by 1.5–2.5 times. Cache went up the most, and a double multiplier applies during peak hours.
Sber launches GigaAgent: an agent pilot inside the Russian perimeter
Sber has opened access to an autonomous general-purpose AI agent. The launch runs through Cloud.ru Agents Space, with a RUB 4,000 starter grant for new users.
OpenAI Cuts Luna Pricing by 80%: Volume Workloads Recalculated
Input on GPT-5.6 Luna dropped from $1 to $0.20 per million tokens; on Terra, from $2.50 to $2. Quarterly budgets are being redone.
FSTEC order on AI: what to prepare before September 1
FSTEC has published a draft order to replace No. 21. For the first time, information security in AI use is set out as a separate area, with the rules taking effect on September 1, 2026.
Anthropic Ships Claude Opus 5: Nearly Fable at Half the Price
The model comes close to flagship Fable 5 at half the price and becomes the default workhorse for business use.
AI law signed: regulatory threshold set at 1 billion parameters
Russia signed its first standalone federal AI law on 26 July. It introduces a threshold of 1 billion parameters and requires companies to appoint someone accountable for compliance.
Run the numbers
How much a move from the current model to a cheaper one costs, and how many weeks the token price difference needs to repay the investment in re-testing prompts.
How many messages, conversations and document pages fit into the window of a chosen model, and at which message the history starts getting evicted.