Four things moved this week that a solo founder should actually price into what they ship — and one throughline they add up to. All verified, all sourced, each with the one line that matters for a team of one.

1. GPT-5.6 is fully GA — the three-tier menu is set#

OpenAI finished the GPT-5.6 rollout on July 9. It's now generally available across ChatGPT, Codex, and the API, and the pricing is confirmed, not rumored. Per million tokens: Sol $5 in / $30 out, Terra $2.50 / $15, Luna $1 / $6. The three names are what OpenAI calls "durable capability tiers" — each can advance on its own cadence, so you're routing to a tier, not a version number.

The benchmark that decides your bill is Terminal-Bench 2.1: Terra ties Claude Fable 5 at 84.3%, while Sol scores 88.8% and Luna 82.5%. Terra sits within two to three points of Sol on most agent evals at half the price.

What it means: "Default to Terra" is now the honest rule. Route your everyday agent loop to Terra, reserve Sol for the hardest planning-and-coding runs, and push high-volume, latency-sensitive calls to Luna. If you hard-coded a model string during the government-gated preview, this is the week to swap it for a tier and re-check your per-task cost. We broke the three tiers down in Sol vs Terra vs Luna, and put Terra head-to-head with the cheap open models in which cheap agent model to route to.

2. China's persona law lands July 15 — Doubao and Qwen pull their agents#

The Interim Measures for the Administration of AI Anthropomorphic Interactive Services — co-issued in April by the Cyberspace Administration of China and four partner agencies — take effect tomorrow, July 15. Neither ByteDance's Doubao nor Alibaba's Qwen could get their consumer custom-agent architecture compliant in time, so both are shutting those features down.

The data handling diverges sharply. Doubao gives users read-only access to their agent configs and chat histories until October 15, then processes the data per its privacy policy — and redirects them to Maoxiang, a separate ByteDance app built for compliance. Qwen has announced permanent deletion with no grace period and no migration path.

What it means: This is a portability fire drill you should run on your own stack today. If a platform can delete your users' agent memory on a regulator's timeline, "the model owns the memory" is a liability, not a feature. Export paths and user-owned memory stores stop being nice-to-haves the moment a jurisdiction you sell into writes a rule. We covered the shutdown mechanics in the Doubao and Qwen agent shutdown and the founder checklist in what founders should do now that it's in effect.

3. Claude Sonnet 5 is the new default — near-Opus at Sonnet price#

Anthropic shipped Claude Sonnet 5 on June 30 and made it the default for every Free and Pro user on claude.ai the same day. It posts 63.2% on SWE-bench Pro and 80.4% on Terminal-Bench 2.1 — the latter actually beating Opus 4.8's 74.6% — and it's priced at an introductory $2 / $10 through August 31 before moving to $3 / $15.

What it means: Between Sonnet 5 and GPT-5.6 Terra, the mid-tier is where the frontier moved this quarter. If your agent's default model is still a flagship from Q1, you're paying a premium for capability your task no longer needs. Re-benchmark your own workload against both mid-tiers before the introductory windows close — we ran the numbers for a bootstrapped stack in Sonnet 5: cheaper agents for founders.

4. Microsoft Agent Framework 1.11 — your tool catalog stops taxing every turn#

Microsoft Agent Framework shipped 1.11, and the headline for anyone running a large tool set is progressive MCP tool discovery: agents can now discover, load, and unload MCP tool schemas on demand within a single run, while the allowed_tools boundary stays intact. Tools can be added or removed based on prior tool results, mid-run.

What it means: Every tool schema you expose is tokens the model re-reads on every turn. Progressive disclosure lets you register a hundred tools and only pay for the handful in play right now — the same lever that makes big MCP catalogs affordable instead of context-bloating. If your agent's system prompt is fat with tool definitions, this is the pattern to copy even if you're not on MAF. We walked through the framework's implementation in progressive MCP disclosure, and the complementary token-saver in how to cache your agent's tool definitions.

5. The quieter throughline: the model price war is now the founder's advantage#

Three of this week's five items are prices falling. Terra at $2.50 input ties a frontier coding model from three months ago. Sonnet 5 undercuts its own predecessor. The tool-schema tax is getting engineered away. For a bootstrapped builder, the compounding effect is that the cost of a capable agent keeps dropping while you sleep — which means the discipline that pays is re-pricing your stack every few weeks, not picking one model and forgetting it.

Do this before Friday: swap any hard-coded model string for a routing tier, run one real task through Terra and Sonnet 5 side by side, and add an export path to wherever your agents keep memory. Three small moves; all three got cheaper or more urgent this week.