JULY 13, 2026 — THREE DECISIONS THIS WEEK
After the most active AI launch week since GPT-4, this week is about what comes next and what needs to be done now. Three stories, each with a time-sensitive action item.
- Gemini 3.5 Pro Targets Thursday July 17 — 2M token context (largest by 2x), Deep Think reasoning (Ultra only), ~$15/$60/M. Only major frontier model with no government restriction. Four senior Gemini researchers left for Anthropic. Action: prepare your benchmark workloads for Thursday's launch. Full preview →
- China Considering Overseas AI Restrictions — Anthropomorphic Rules Take Effect Today — Reuters: Chinese authorities met Alibaba, ByteDance, Z.ai about restricting overseas model access. No official announcement. Would reverse China's open-weight strategy. Also today: anthropomorphic AI enforcement window opens. Action: start contingency planning if you use Chinese AI APIs for production. Full story →
- DeepSeek API Migration — July 24 Hard Deadline — 11 days. deepseek-chat and deepseek-reasoner both break. Migrate to deepseek-v4-pro or deepseek-v4-flash. Critical: deepseek-reasoner maps to V4-Flash, not V4-Pro — heavy reasoning users must call V4-Pro explicitly or lose capability. Also: Mistral Leanstral 1.5 generates Lean 4 mathematical proofs that software behaves as intended. Full migration guide →
Story 1 — Gemini 3.5 Pro: The Model Everyone Has Been Waiting For Targets Thursday
Google is targeting July 17 for Gemini 3.5 Pro's general availability, according to Business Insider — four weeks after missing its June deadline. Gemini 3.5 Pro is the only major frontier model without government access restrictions. A broadly available Pro model beats a benchmark-topping model that most people cannot access yet. GPT-5.6 Sol just cleared a 13-day government review. Grok 4.5 is unavailable in the EU. Claude Fable 5 requires credits. Gemini 3.5 Pro's availability advantage is real and immediate.
Reported specs: 2 million token context window — the largest in any production model by a factor of two — Deep Think extended-inference reasoning restricted to $250/month Ultra subscribers, and estimated pricing of approximately $15/$60 per million tokens. The SWE-bench Pro score has not been published — that is the benchmark most relevant for developer adoption, and it will be the most-watched number when the model card drops on Thursday. The competitive front has moved to who can be trusted to run a 30-step task unsupervised. That is the bar Gemini 3.5 Pro has to clear. Read the full Gemini 3.5 Pro preview →
Story 2 — China's Overseas AI Restrictions: What Would Change and What It Means Today
On July 7, 2026, Reuters reported exclusively that Chinese authorities held meetings with Alibaba, ByteDance, and Z.ai about potentially restricting overseas access to China's most advanced AI models. No comment from the Ministry of Commerce, NDRC, Alibaba, ByteDance, or Z.ai. This is an unconfirmed report at the meeting stage — not a policy announcement. But it deserves developer attention because the symmetric logic is clear: the US restricted Fable 5 and Mythos 5 on cybersecurity grounds, and China faces the same capability concerns in reverse.
The strategic stakes: China's open-weight strategy worked — GLM-5.2, Qwen, and DeepSeek drove global goodwill and developer lock-in. Retreating now wastes leverage against US closed APIs. Restricting overseas access would hand Western labs their largest competitive opening since 2023. Teams relying on DeepSeek V4, Qwen, or GLM APIs for production should begin contingency planning — not urgent migration, but a documented alternative with estimated migration cost. Separate and urgent today: China's anthropomorphic AI rules enforcement window is open. ByteDance and Alibaba have complied. Third-party apps that have not updated face removal from Chinese app stores this week. Read the full China AI restrictions story →
Story 3 — DeepSeek July 24 Deadline: The Migration Error That Will Break Pipelines
Every development team using DeepSeek's hosted API has one non-optional task: update any code that calls deepseek-chat or deepseek-reasoner before July 24, 2026 at 15:59 UTC. Both aliases stop working with no announced extension. The mechanical change is trivial — one parameter. The risk is in the alias mapping.
deepseek-reasoner maps to V4-Flash (thinking mode), not V4-Pro. If you were using deepseek-reasoner for heavy reasoning tasks and need equivalent capability, you must explicitly call deepseek-v4-pro, not just replace the alias name. Teams that replace deepseek-reasoner with deepseek-v4-flash assuming parity will silently degrade their reasoning pipeline quality. Test before July 20 to leave four days of buffer. Also worth watching: Mistral launched Leanstral 1.5 — a 119B Apache 2.0 model that moves beyond code generation to provide mathematical proof that software behaves as intended using Lean 4, potentially transforming reliability verification for critical systems. Read the full migration guide with alias mapping table →
This Week's Action Items
TODAY — Audit all DeepSeek API calls. Search your repos for deepseek-chat and deepseek-reasoner. 11 days to the July 24 hard deadline. The migration is trivial; finding every call site is the work.
THURSDAY — Gemini 3.5 Pro launches. Prepare your benchmark workloads. SWE-bench Pro, long-context recall validation, and agentic tool-use tests are the three evaluations that will determine whether 3.5 Pro changes your stack. Have your test suite ready to run the moment the API is accessible.
THIS WEEK — China overseas restriction contingency. If your production stack includes DeepSeek V4, Qwen, or GLM hosted APIs, document a migration path to an alternative now. Not urgent deployment — a written plan with estimated migration cost. If restrictions are formalised, you will want 30 days to migrate, not 30 days to plan.
BEFORE AUGUST 31 — Benchmark Claude Sonnet 5 tokenizer costs. Intro pricing ($2/$10/M) ends August 31. Terra ($2.50/$15/M) and Gemini 3.5 Pro (~$15/$60/M) are now live comparators. The tokenizer multiplier (1.0-1.35x) means Sonnet 5 real costs are $2.60-$3.90/M input. Benchmark before September 1 price step-up.