Summary

The late-July frontier race sharpened as Anthropic launched Claude Opus 5 — 1M-token context, a new five-level effort dial, and 'frontier at half the price of Fable 5' held at Opus 4.8's $5/$25 rate — one day after DeepSeek moved its V4 family (V4-Pro and V4-Flash, both 1M context) into general availability and hard-retired its legacy deepseek-chat and deepseek-reasoner aliases with new peak/off-peak pricing. Underneath the model launches, the infrastructure to run agents kept maturing: Google expanded Gemini API Managed Agents with long-running background tasks, remote MCP server integration, and custom function execution, pushing its hosted runtime toward parity with AWS Bedrock AgentCore and Anthropic's Managed Agents. And the physical layer scaled with it, as NAVER, NVIDIA and Brookfield unveiled a ~$10B plan to grow Korea's national AI factory from 55MW to 200MW on Vera Rubin and Blackwell systems.

Key themes

  • Frontier price-performance compression: Claude Opus 5 and DeepSeek V4 GA landed within a day of each other, both leaning on cost controls (Opus 5's five-level effort dial, DeepSeek's peak/off-peak pricing) in a window that also includes GPT-5.6 and Kimi K3.
  • Managed-agent runtimes and MCP hardening: Google's Gemini Managed Agents update adds background tasks and remote MCP servers, signaling MCP's consolidation as the default tool-calling standard across every major hosted-agent platform.
  • Sovereign AI infrastructure and its financing: the NAVER/NVIDIA/Brookfield 200MW buildout advances the national-AI-factory pattern, with an infrastructure investor (Brookfield, up to $9B) underwriting most of the capex rather than the operator's balance sheet.
  • Migration pressure on production teams: DeepSeek's hard cutover of legacy API aliases forces developers off hard-coded endpoints, a reminder that GA milestones increasingly come with firm deprecation deadlines.

Notable items

  • Anthropic launched Claude Opus 5 (July 24) with a 1M-token context window, up to 128k output tokens, a new five-level reasoning effort control, and flat $5/$25 per-million pricing (same as Opus 4.8, plus a ~2.5x Fast mode at $10/$50), on the Anthropic API and Amazon Bedrock — the day's only high-impact signal.
  • DeepSeek V4 reached general availability (July 20), shipping V4-Pro (1.6T total / 49B active) and V4-Flash (284B / 13B active) with 1M-token context and agentic gains; legacy deepseek-chat and deepseek-reasoner aliases went inaccessible after July 24 15:59 UTC, alongside new peak/off-peak API pricing.
  • Google expanded Gemini API Managed Agents (announced July 7) with long-running background task support, remote MCP server integration, custom function execution, and streamlined credential management, moving the Google-hosted runtime from I/O preview toward production parity with Bedrock AgentCore and Anthropic Managed Agents.
  • NAVER, NVIDIA and Brookfield announced a ~$10B plan (July 25) to expand Korea's national AI factory at GAK Sejong from 55MW to 200MW by 2028 on NVIDIA Vera Rubin and Blackwell platforms — Brookfield funding up to $9B, NVIDIA $1B — as part of NAVER's path toward 1GW of NVIDIA infrastructure.

Source coverage

Source rows used: 4