Yesterday's releases tightened the cost curve on every tier of the frontier. Anthropic's Opus 5 makes near-flagship performance accessible at subscription prices last seen in mid-range models. DeepSeek V4's general availability adds a 57x price differential at comparable output quality. Two days from now, Kimi K3 open weights will make self-hosted frontier-class coding possible for the first time. The EU's DMA enforcement, meanwhile, is reshaping who can reach two billion Android devices with AI.
Anthropic Ships Opus 5: Effort Toggle Reshapes Enterprise Cost Calculus
Anthropic released Claude Opus 5 on 24 July, its fourth model in under two months. Priced at $5 per million input tokens and $25 per million output tokens — matching Opus 4.8, and exactly half of Fable 5 — it ships with a 1-million-token context window and is available immediately on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, claude.ai, Claude Code, and Cowork.
On CursorBench 3.2, Opus 5 at maximum effort finishes within 0.5 percent of Fable 5's peak score on coding tasks while cutting cost per task by approximately half. On hard-coding evaluations it roughly doubles Opus 4.8. The model sits behind Mythos 5 on offensive cybersecurity benchmarks — a design choice, not a capability gap.
The most operationally significant feature is the per-request effort toggle (low, medium, high). Until now, cost management meant choosing a weaker model or accepting higher bills. An effort level allows cost-sensitive pipelines — batch summarisation, document triage, classification — to share infrastructure with high-stakes agentic tasks running at full effort, without changing model identifiers or splitting deployment environments. That is the pricing structure enterprise procurement teams should be stress-testing against current Fable 5 and Opus 4.8 contracts.
EU Fines Google €890 Million and Gives It 60 Days to Redesign Search
The European Commission on 23 July imposed a €890 million penalty on Alphabet — its first under the Digital Markets Act and the largest DMA fine to date. The ruling covers two distinct violations: favouring Google Flights, Hotels, and Shopping in search results over rival services (€460 million); and restricting Play Store developers from directing users to cheaper off-platform purchase channels (€430 million). Google's cumulative EU liabilities now exceed €10 billion.
The 60-day deadline to redesign Search is the active lever. Non-compliance triggers periodic fines of up to 5 percent of Alphabet's average daily global turnover. Separately, binding interoperability orders issued on 16 July require Google to open 11 core Android features to competing AI assistants by July 2027 — voice wake-word activation, in-app action capabilities, suggested replies, and contextual queries — and to share anonymised Search data with rival AI and search providers from January 2027.
For operators managing European enterprise fleets on Android, this creates a credible 12-month window for non-Gemini AI assistants to reach equivalent system-level access. The Search data-sharing provision is a structural shift for competitors: it increases training-data availability for Anthropic, OpenAI, and European AI developers in ways that compound over model generations.
Kimi K3 Open Weights Arrive Sunday: 2.8 Trillion Parameters, 1.4 TB Floor
Moonshot AI confirmed that Kimi K3 weights will publish on Hugging Face on 27 July under a Modified MIT licence. K3 is a 2.8-trillion-parameter Mixture-of-Experts model with 16 of 896 experts active per token — approximately 50 billion active parameters — giving it per-token compute closer to a mid-size model. Kimi Delta Attention delivers up to 6.3 times faster decoding than standard attention at comparable quality.
The infrastructure requirement is the operative constraint. At MXFP4 four-bit precision the weights need approximately 1.4 terabytes of fast memory; at 16-bit, 5.6 TB. That places self-hosting in cloud bare-metal or dedicated HPC territory, not standard GPU cluster configurations. Independent testing flagged a 51 percent hallucination rate on factual recall queries — a figure Moonshot omitted from its published benchmark suite and one that requires mitigation design in production deployments.
Via API, K3 sits second overall on the Vals AI index (behind Fable 5 and GPT-5.6 Sol Max) and holds first place on the Frontend Code Arena. For organisations with the infrastructure to self-host — and a tolerance for the hallucination caveat — this is the first open model genuinely competitive with frontier closed models on coding and agent tasks. The Modified MIT licence permits commercial use and fine-tuning without royalties.
DeepSeek V4 Is Live: Legacy Endpoints Went Dark Yesterday
DeepSeek V4 reached stable general availability on 24 July, graduating from its April preview. The legacy deepseek-chat and deepseek-reasoner API aliases were decommissioned at 15:59 UTC on 24 July. Any production integration that has not migrated to deepseek-v4-flash or deepseek-v4-pro will be returning errors as of today.
V4 introduces time-variable pricing: API costs double during Beijing business hours (09:00–12:00 and 14:00–18:00 CST). Workloads that run continuously need explicit peak-hour cost modelling; workloads that can be scheduled outside those windows see the headline off-peak rate. Early benchmarks place V4-Pro's performance close to Opus 4.8 on general tasks, approaching GPT-5.6 Sol on code, at approximately 1/57th of Fable 5's API price.
The migration is mechanical but time-sensitive: update the model field to deepseek-v4-flash for standard tasks or deepseek-v4-pro for heavier reasoning, confirm the endpoint base URL, and validate that peak-pricing windows are reflected in cost projections. Both new model IDs are exposed via OpenAI-compatible and Anthropic-format DeepSeek endpoints.
The through-line across this week: AI procurement benchmarks set six months ago are structurally out of date. Opus 5 halves effective near-frontier cost. DeepSeek V4 offers Opus-class output at a fraction of that. Kimi K3 brings self-hosted frontier-class coding within reach for the first time. Operators still sizing budgets against Fable 5 or GPT-5.6 Sol pricing are overestimating costs for most workloads — and potentially underestimating what is now available at the lower tiers.