Uncle Sam cleared GPT-5.6 (and Elon is capping our budgets)

Another chaotic week of shipping code and paying API bills. Between OpenAI getting Washington's permission to release GPT-5.6 and Tesla slamming the brakes on token spending, our development stacks are getting highly volatile. Here's what actually matters to those of us trying to ship code with these models. ### This Week in AI Coding * **GPT-5.6 Sol, Terra, and Luna are live.** OpenAI publicly rolled out its new model family on July 9, after a two-week gated preview tied to a US-government safety review. Sol is the flagship for hard reasoning ($5/M input, $30/M output), Terra is the everyday-use tier ($2.50/$15), and Luna is the budget option ($1/$6) — all three carry a 1.05M context window. **My take:** I'm routing my recursive test loops to Luna; at a buck a million input tokens, my automated agents can run wild without triggering a credit-limit panic. * **Anthropic ships a "Reflect" dashboard.** Rolled out July 9 for Free, Pro, and Max users with memory enabled — it tracks and visualizes your Claude usage over 1/3/6/12-month windows, offers quiet hours, and nudges you with questions like "what's one thing you want to keep doing yourself, even if Claude could do it faster?" based on Anthropic's own 4D AI Fluency framework. **My take:** I don't need my coding assistant asking me deep questions mid-deploy, but the nudge toward using Projects instead of re-pasting context every session will actually save you tokens. * **The "Friendly Fire" agent exploit.** AI Now Institute published a proof-of-concept (July 8) showing Claude Code in auto-mode and Codex in auto-review can be tricked into executing malicious code just by being pointed at a repo and asked to "run a security review." The payload disguises a malicious binary next to a decoy Go source file so it looks like something the agent legitimately compiled — and the same trick worked unchanged across Sonnet 4.6, Sonnet 5, Opus 4.8, and GPT-5.5. **My take:** if you're letting an agent run "auto" security scans on unvetted public repos, you're one README away from handing over your host. Keep command-capable agents away from secrets without a human in the loop. * **Tesla caps AI spending at $200/week per employee**, effective July 6 — engineers had been burning "thousands of dollars" in tokens weekly. The cap conveniently exempts xAI's beta products (including Grok), so heavy users get pushed toward Musk's own models instead of Claude or GPT. **My take:** gamifying token usage for months and then slamming the brakes — while carving out your own AI company — is a very on-brand way to lose developer goodwill. * **Gemini 3.5 Pro delayed to July 17** for a full architectural rebuild — Google DeepMind scrapped the 2.5 Pro base entirely rather than ship an incremental update. The rebuilt model targets a 2M-token context window and a "Deep Think" reasoning layer aimed squarely at GPT-5.6 and Fable 5. **My take:** tearing up your architecture weeks before launch is a costly move — it only makes sense if Google genuinely believes incremental won't cut it against this quarter's frontier releases. * **Kimi K2.7 Code is now GA in GitHub Copilot** (July 7) — the first open-weight model in Copilot's model picker, rolling out to Pro/Pro+/Max now, with Business/Enterprise admins needing to flip on a policy first. **My take:** having a real open-weight option next to the closed models in Copilot's picker is a good hedge against vendor lock-in, especially if your legal team gets nervous about closed-weight terms. ### What We Shipped We had some tough conversations at VeriduxLabs this week and made some necessary cuts: * **Retired SPARit (sparit.vc).** After ~2 months as our lead launch product — ads in English and Hebrew, Shorts, build-in-public threads, a daily acquisition loop — it landed 0 paying customers and no funnel conversion. Both Google Ads campaigns are paused, the acquisition routines are disabled, and the site stays up as an internal tool only. A clean kill, documented publicly. * **Consolidated focus to a few directions instead of many.** Primary bet now: **TypeWright** (typewright.dev) — a VS Code fork running a local 1.5B TypeScript model with a QLoRA fine-tuning pipeline and an Ed25519 licensing server, being built toward an exit with a live catalyst window around July 12. * The surviving stock-analysis work continues only in the research track (champion/challenger evolutionary protocol), not as a shipped product. * No new YouTube uploads this week on @endofcoding — the backlog was already cleared. ### Tool of the Week **Cursor 3.11** (shipped July 10) — the smaller update that makes Cursor's agent-first "Glass" interface actually livable day to day. Side Chats (`/side`, `/btw`) let you spin up parallel, durable agent conversations without losing your main thread. Conversation Search finally makes past agent transcripts searchable via Cmd+K/Cmd+F across thousands of sessions. New Cloud Agent hooks (`beforeSubmitPrompt`, `afterAgentThought`, `subagentStart`, etc.) let you observe and steer subagents programmatically. If you're already living in Cursor's agent windows, this is the release that stops you from losing context every time you branch off to check something. ### Quick Links * [Anthropic: Introducing Reflect](https://www.anthropic.com/news/reflect-with-claude) — the official writeup on Claude's new usage-reflection dashboard. * [AI Now Institute: Friendly Fire exploit brief](https://ainowinstitute.org/publications/friendly-fire-exploit-brief) — the full research behind this week's agent-hijacking finding. * [GitHub Changelog: Kimi K2.7 Code for Copilot Business/Enterprise](https://github.blog/changelog/2026-07-07-kimi-k2-7-now-available-for-copilot-business-and-enterprise/) — rollout details and how to enable the policy. * [Cursor changelog](https://cursor.com/changelog) — full 3.11 release notes, including the new cloud agent hooks. --- Until next week — keep shipping, and watch those spending caps. Guy [endofcoding.com](https://endofcoding.com) | [vibecodingebook.com](https://vibecodingebook.com) | [llmhire.com](https://llmhire.com) | [typewright.dev](https://typewright.dev) | YouTube [@endofcoding](https://youtube.com/@endofcoding)