The short version: The model launches got the headlines this week; we covered the GPT-5.6 GA and the macro picture in the July 14 wire. This is the other half — the developer diff. Between July 7 and 12, the SDKs underneath your agent shipped real releases, and at least one of them can change which model your app runs on the next time you run pip install -U. Every line below is sourced to a GitHub release note, not an aggregator, because these are the ones you act on. Here's the upgrade checklist.

1. OpenAI moved the default model — check your routing#

Two OpenAI packages shipped. openai-python v2.45.0 (July 9) added GPT-5.6 support. More consequentially, openai-agents-python v0.18.0 through v0.18.2 (July 7–11) changed the default model to GPT-5.6, added GPT-5.6 request controls, and — per the changelog — added "hosted multi-agent beta support."

The action item is unglamorous but real: an upgrade can silently move which model tier your agents call. If your app relied on the previous default, pin your model explicitly rather than inheriting whatever the SDK now points at. And if you're orchestrating multiple agents on OpenAI, the hosted multi-agent beta is worth a look — it's the SDK offering to run the loop server-side instead of you hosting it, which reframes the build-vs-buy question on your orchestration layer. (The changelog names the capability; treat the exact API surface as beta and read the release notes before you lean on it.)

2. Microsoft Agent Framework 1.11 — stop paying tool tax every turn#

microsoft/agent-framework python-1.11.0 (July 10) is the most directly cost-saving release of the week. Its headline is progressive MCP tool disclosure: instead of loading every connected tool's full schema into context on every turn, the agent discovers tool schemas on demand. It also added message-injection middleware, skills caching with a TTL, and an approval opt-out for skills and file tools.

If your agent is wired to a dozen MCP servers, you have been spending tokens every single turn describing tools it never calls. Progressive disclosure is the framework finally letting you stop.

This is the same instinct behind MCP's own move to trim what servers push at clients: the tool catalog shouldn't tax every request. If you run a tool-heavy agent over MCP, this is the upgrade to schedule.

3. LiteLLM day-0'd Gemini 3.5 Flash — a cheaper fast tier, already routable#

Gateways are where "a new model launched" becomes "a new model is callable in your code without a rewrite." LiteLLM v1.92.0 (July 12) shipped day-0 support — cost map plus provider wiring — for Gemini 3.5 Flash, alongside cost-map entries for Claude Opus 4.8 and Fable 5. For a high-volume, latency-sensitive workload, a new cheap-fast Gemini tier you can route to behind an existing gateway call is exactly the kind of quiet cost lever that compounds.

One caution: the release notes referenced the Gemini SKU inconsistently, so confirm the exact model ID against Google's own model list before you wire it into production routing. Day-0 support means the plumbing is there; it doesn't mean the string in a third-party changelog is the one you should ship. This is the ongoing payoff of putting a gateway between your app and the model providers: you adopt a new model by changing a string, not a dependency tree.

4. The dated one: Sonnet 5's promo pricing expires August 31#

Not a release, but the item with a deadline. Claude Sonnet 5's introductory pricing — $2 per million input tokens, $10 per million output — is promotional through August 31, 2026, confirmed in LiteLLM's model-cost file and the Claude Code changelog. Reporting points to a step-up after the promo window. If you sized an agentic budget — where output tokens dominate — on the promo rate, model the post-August cost now while you can still choose. Sonnet 5 as the near-Opus default at Sonnet price is a genuinely good deal; just know the deal has an end date on the calendar.

What we're watching but won't call yet#

Two items are real but rest on secondary sourcing this week, so we flag them rather than bank them. xAI's Grok 4.5 — an "Opus-class" coding model reportedly trained alongside Cursor and available in Cursor across plans — is corroborated by TechCrunch and appears in gateway cost configs, but xAI's own announcement was not reachable to verify first-hand; it's worth A/B-testing against your current coding model if you live in Cursor. And Prime Intellect's $130M Series A for decentralized/open-model training points at more open weights for self-hosters, but the number is aggregator-sourced for now.

The throughline is the one that keeps repeating for solo builders: the cost of the same behavior keeps falling, and the friction to adopt the cheaper option keeps shrinking — a gateway day-0's the model, a framework stops taxing your tool catalog, a promo rate resets your unit economics. The tax is that your defaults now move under you between releases. Read the diff, pin what you depend on, and take the cuts.