This week produced one direct API price improvement and several changes to the less-visible parts of the AI bill. Google launched Gemini 3.6 Flash with lower output pricing than its predecessor. xAI opened Grok 4.5 to EU developers. Anthropic changed how Fable 5 consumes paid-plan usage, while OpenAI’s Codex context adjustment and Cursor’s agent-swarm analysis put attention on context, parallelism, and quotas.
| Story | What changed | Buyer impact |
|---|---|---|
| Gemini 3.6 Flash | New GA model lowers output pricing versus Gemini 3.5 Flash | Output-heavy agents get the clearest rate-card saving |
| Grok 4.5 in the EU | EU users can access the model in xAI’s API console | More frontier choice, but no announced discount |
| Claude Fable 5 plans | Fable is available across paid plans with plan-specific usage treatment | Subscription buyers must track quota burn, not only access |
| Codex context | OpenAI’s Codex repository reduced several model context settings from 372k to 272k | Large-repository workflows may need tighter context selection |
| Agent swarms and local models | Parallel agents and local frontier models gained attention | Effective cost now depends heavily on concurrency and routing |
Gemini delivered the week’s direct price change
Gemini 3.6 Flash and Gemini 3.5 Flash-Lite reached general availability. Google kept Gemini 3.6 Flash’s input rate aligned with the prior Flash generation while lowering its output rate. That makes the new model most attractive for coding, tool use, and reasoning loops that produce substantial output.
The migration is not only a model-name change. Google says temperature, top_p, and top_k are deprecated and ignored on the latest models, and future generations will reject them. Teams should remove those fields and prefilled model turns before comparing quality or cost.
Read the full Gemini 3.6 Flash pricing and migration guide and verify current rates on the Google AI pricing page.
Grok 4.5 became available to EU developers
xAI says Grok 4.5 is now available in its API console for EU users. This expands access; it does not change the rate card, consumer subscriptions, or guarantee that inference data stays in the EU.
European teams should confirm processing location and retention terms, pin the exact model identifier, and benchmark successful-task cost at each reasoning setting. Our Grok 4.5 EU launch analysis covers the rollout, while the xAI pricing page supplies live rates.
Claude Fable 5 changed the subscription calculation
Anthropic’s updated help page says Fable 5 is available on all paid Claude plans. Max subscribers can use it for up to half of their regular weekly usage limit at no extra cost. On Pro and standard Team or legacy Enterprise seats, Fable use draws on pay-as-you-go credits.
The practical comparison is therefore plan allowance versus API usage, not “included” versus “paid.” Heavy Fable users should record how quickly the model consumes weekly quota and what happens after the included allocation is exhausted. Compare current API options on the Anthropic pricing page.
Context and parallel agents became cost controls
An OpenAI Codex repository change reduced several model context settings from 372k to 272k tokens. That is not an API price increase, but it can change repository indexing, compaction, and prompt-selection behavior for large coding tasks.
Cursor separately argued that agent swarms change model economics: cheap parallel attempts can outperform one expensive sequential route, but only when coordination overhead and duplicate work stay controlled. Teams should measure total tokens across every branch, wall-clock time, merged output, and human cleanup.
Open and local routes kept pressure on hosted pricing
Kimi K3 demand reportedly became strong enough for Moonshot AI to pause new subscriptions, while projects such as Nativ and Petals highlighted local or distributed model execution. Demand validates the open-model market, but capacity limits also show that low list prices do not guarantee immediate access.
For teams that want managed open-model access without operating GPUs, Novita is one OpenAI-compatible option to benchmark against first-party and local routes.
Affiliate disclosure: this article contains an affiliate link. It may earn us a commission at no extra cost to you and does not change the analysis.
What buyers should do now
Move Gemini Flash evaluations to the new GA models after cleaning up deprecated request fields. Add Grok 4.5 to EU test matrices, but verify data terms before regulated use. For Claude subscriptions, monitor quota consumption by model instead of assuming all included usage has equal value.
For coding agents, track context loaded, compactions, parallel branches, failed attempts, and accepted changes. The token calculator can model the rate-card layer; your own task telemetry must supply the retry and success layer.
Sources: Google’s latest Gemini models guide, xAI’s release notes, Anthropic’s Fable 5 plan guide, the OpenAI Codex context-window change, Cursor’s agent-swarm economics analysis, and AI Pricing Guru’s live pricing dataset.