AI news June 28 2026: GPT-5.6, Gemini browser control, SpaceX at $6B
Alexandre
··
Reading time: 9 min
Click to enlarge
OpenAI ships GPT-5.6 at half the price of Anthropic. Sol, the flagship, bills at $5 per million input tokens versus $10 for Fable 5. Access is restricted to roughly 20 companies approved by Washington, but the pricing is public and the message is clear.
Other numbers landed this week: Colossus, SpaceX's Memphis datacenter, is pulling in $1.25 billion per month from Anthropic and $920 million from Google. DeepSeek raises $7.4 billion at a $50 billion valuation and announces it's doubling headcount. Gemini 3.5 Flash can now control a browser in production via the Google API. Karpathy formalizes ten rules for autonomous code loops.
What this week says about the structure of the AI ecosystem: models proliferate, prices fall, but physical infrastructure concentrates. I build Waku with Claude Code, and the question of who controls the GPUs affects me directly.
Let's break it down.
GPT-5.6: Sol, Terra, Luna land at half price, under Washington's watch
OpenAI launches on June 26 a suite of three models: Sol (flagship), Terra (high-volume use) and Luna (fast and lean). Sol bills at $5 per million input tokens and $30 per million output, roughly half the price of Claude Fable 5 ($10 / $50). Terra costs half of Sol, Luna half of Terra. Access restricted to roughly 20 companies approved by the Trump administration, according to The Verge.
Sol introduces two new modes: "max" for deep reasoning, and "ultra" which deploys coordinated sub-agents for long and complex tasks. OpenAI positions it as particularly strong in code, cybersecurity, and biology. On published benchmarks, Sol claims performance above Mythos on GeneBench, Terminal-Bench, and ExploitBench, with about a third fewer tokens consumed. OpenAI also improves prompt caching, reducing costs on repetitive requests.
Sam Altman accepted the government limitation without much pushback. OpenAI stated it was "complying with the government's request on an exceptional basis," while noting that "restrictions shouldn't become the norm." Diplomatic messaging, but the signal is clear: shipping frontier models is now a process that involves Washington.
Pricing is the real story. 5/30 versus 10/50 for Fable 5 is an open price war. Anthropic will have to respond. For a dev running agents in production, the cost difference shows up very fast on the monthly bill. OpenAI also announces Codex usage is up: 5x more active users in six months, and adoption multiplied by 137 among non-developers. These numbers are not neutral: they say AI coding tools have left the dev niche.
But while models proliferate, the tools to control them are evolving too. And Google just crossed a threshold.
Click to enlarge
Gemini 3.5 Flash takes control of your browser
Google announces on June 24 that Gemini 3.5 Flash now integrates a native computer-use tool, replacing the dedicated Gemini 2.5 Computer Use model. The capability allows agents to "see, reason, and act" in browser, mobile, and desktop environments: continuous software testing, automation of complex workflows on professional applications, according to 9to5Google. Available now via the Gemini API and the Gemini Enterprise Agent platform.
Computer-use is the ability for an AI model to visually perceive a screen and control it the way a human user would: click, type, navigate, fill forms. This is no longer a lab prototype, it's now a production tool in Google's API. Gemini in Chrome simultaneously adds a "Select from screen" tool that lets you zoom in on a specific area to contextualize a prompt.
Guardrails exist: explicit user confirmation for sensitive or irreversible actions, automatic stop if an indirect prompt injection (malicious text that hijacks the AI's behavior) is detected. Google hosts a test environment via Browserbase.
Google is also delaying Gemini 3.5 Pro to July. The model is available to testers on the Antigravity platform, but the company wants more field feedback before the public release, according to Business Insider.
To me, generalist computer-use in production is the real signal from Google this week. Not the model itself, but what it enables: an agent that can navigate any interface without a dedicated API. It's a direct attack against all the integrations that assumed AI needed a structured API to interact with a system. For solo devs building automation agents, it opens use cases that were impossible cleanly six months ago.
And while models learn to control interfaces, another shift is underway on the tooling side.
CLAUDE.md grows to ten rules: loop engineering becomes a discipline
Karpathy and the Claude Code team formalize this week a ten-rule protocol for autonomous code loops. The /goal command, added in Claude Code 2.1.139, triggers automatic stop when a verifier model confirms the objective is reached, with no human intervention at each turn, according to Tech Times.
Loop engineering is the practice of building systems that prompt an AI agent, evaluate its output against explicit success criteria, and repeat until a test passes. The new protocol names four failure modes to avoid: Kitchen Sink (too many problems at once), Wrong Abstraction (bad problem representation), Optimistic Path (assuming the happy path without handling errors), Runaway Refactor (refactoring that never converges).
The central rule: the model that writes the code doesn't verify its own work. A separate, faster verifier model takes that role. Clean separation that avoids self-validation bias. Boris Cherny, Claude Code's creator, frames the problem clearly: autonomous loops can cost up to fifteen times the price of a normal conversation when they spiral without supervision, reports Business Insider.
I've been using Claude Code on Waku for months. Autonomous loops that run expensive, that's a reality I've run into. These ten rules aren't just best practices: they're concrete financial protection for anyone running agents in production. TechCrunch notes that "AI is getting more and more loopy": the ecosystem is converging toward architectures where the agent runs alone, verifies alone, and stops only on an objective criterion. If you build with these tools, reading the new CLAUDE.md before launching your first loop is now essential (I cover it in detail in my article on AI code agent architecture).
SpaceX pulls $1.25B/month from Anthropic: Colossus becomes AI's secret cloud
SpaceX signs on June 22 a compute deal with Reflection AI, a little-known open-source lab: $150 million per month from July 2026 through 2029, meaning $6.3 billion over the term, for access to Nvidia GB300 chips in the Colossus 2 datacenter in Memphis, Tennessee, according to TechCrunch. This deal reveals in passing numbers we had never seen all together.
Here's what Axios and CNBC allow us to piece together on Colossus revenue via published compute contracts:
Anthropic: $1.25 billion per month
Google: $920 million per month
Reflection AI: $150 million per month starting July 1
That's $2.32 billion per month for SpaceX, just from AI compute rental. Colossus 2 in Memphis hosts Nvidia GB300s with immediate access. Reflection AI can terminate with 90 days' notice after an initial three-month period, and SpaceX can too. These are flexible contracts, not locked multi-year commitments.
Reflection AI is an open-source lab founded by former DeepMind researchers, backed by Nvidia. Not a name the general public knows. But signing at $150 million per month to get compute says something about the infrastructure needs of labs building frontier models, even open-weight ones.
This is the revelation of the week for me. We debate models, benchmarks, API pricing. But who owns the GPUs all of this runs on? SpaceX. Elon Musk is renting to Anthropic (his direct competitor via xAI) $1.25 billion per month of compute. He's renting to Google $920 million. The physical infrastructure of frontier AI is centralized in a datacenter owned by a single entity. That's a concentration of power at the physical layer that had never been measured publicly before these contracts.
For me, a solo dev building Waku and using Claude Code daily (I write about this workflow here), it raises a simple question: if Colossus goes down or SpaceX and Anthropic have a dispute, what happens to Claude users? Infrastructure dependency is not an abstract risk.
Click to enlarge
DeepSeek raises $7.4 billion and prepares to double headcount
DeepSeek closes this week its first official fundraising round: 50 billion yuan (roughly $7.4 billion), at a valuation exceeding $50 billion. Tencent, CATL, and China's National AI Investment Fund are among the investors. Founder Liang Wenfeng, who held nearly 90% of the equity before the round, injected roughly $3 billion himself. Immediate stated objective: double the size of every department, according to Investing.com.
27 positions opened simultaneously: dev engineers, data engineers, AI product managers, HR, legal, finance. The expansion also targets a new "Harness" team and explores building proprietary compute infrastructure, for less dependence on Nvidia chips subject to US export restrictions, according to Let's Data Science.
It's the context that makes the number interesting. US-banned Nvidia chips are now selling at double their normal price on China's black market, reports Reuters. DeepSeek doesn't want to depend on them anymore. It's building its own pipeline.
The US administration is simultaneously pressuring Meta to submit to voluntary security reviews on its AI models, according to Reuters. Meta joins OpenAI, Anthropic, DeepMind, Microsoft, and xAI which already participate in these reviews. The subtext: Washington wants to know what all the major labs are doing before models ship.
The American regulation that was supposed to protect a competitive advantage is creating exactly the conditions it was trying to prevent. The 30-day model review before release slows down OpenAI and Anthropic. DeepSeek has no Washington asking it to pause.
Other news
Mythos cracked classified NSA systems: Reuters reveals on June 24 the precise reason for the June 12 Fable 5 ban: a jailbreak had allowed Mythos 5 to find vulnerabilities in classified NSA systems within hours, according to Reuters. The ban itself dates back two weeks, but the public explanation only just dropped. Mythos 5 is partially reinstated Friday for certain US organizations, Fable 5 remains restricted.
OpenAI pushes its IPO to 2027: Sam Altman is leaning toward a listing next year rather than late 2026, according to Forbes. Valuing a company whose models can be banned by government directive makes the exercise complicated.
RAISE US: $500M to retrain AI-displaced workers: OpenAI, Anthropic, Amazon, Microsoft, and Bank of America fund a $500 million joint fund to finance retraining for workers affected by AI. Announced this week with few details on governance.
OpenAI Patch the Planet: bug bounty campaign across open-source software, with GPT-5.5-Cyber reaching 85.6% on the CyberGym benchmark (WIRED). Defensive AI as a counterweight to the crisis of models cracking classified systems.
Polestar banned from US sales: the Chinese-linked automaker is barred from selling in the US under new Chinese technology rules (Bloomberg Law). Export control no longer covers just software.
What this all means for us, indie devs
1. Frontier model prices are going down
GPT-5.6 Sol at $5 input vs Fable 5 at $10 is direct pricing pressure. Anthropic will have to align or justify the gap. For devs in production, it's good news in the medium term, even if the current restricted access limits who can benefit right now.
2. Agents that control interfaces are in production
Gemini 3.5 Flash with computer-use in production API confirms that interface automation without a dedicated API is no longer a prototype. An agent can see a screen and control it. Use cases that assumed a structured API as a prerequisite are worth rethinking.
3. Loop engineering is a skill to acquire now
Ten rules, a separate verifier model, a /goal command: this protocol is actionable immediately. Autonomous loops that cost fifteen times the price of a normal conversation are real. Read the new CLAUDE.md before launching a loop in production.
4. AI's physical infrastructure is more concentrated than we thought
SpaceX's Colossus pulls $1.25B/month from Anthropic and $920M/month from Google. This concentration of physical infrastructure in a single entity is an operational risk for anyone building on frontier models. Diversifying model providers is one thing, but the underlying infrastructure remains a single point of failure.
5. China accelerates while the West regulates
DeepSeek at $50B valuation, $7.4B raised, headcount doubling, building proprietary infrastructure. The US competitive advantage in AI is not a fixed given.
Conclusion: the week physical infrastructure became visible
What stands out about this week is the surfacing of layers we couldn't see. Model pricing is visible. Benchmarks are visible. But who owns the GPUs running Fable 5 and Gemini: that's SpaceX, and we vaguely knew it, but the amounts published this week make the thing concrete and unsettling.
For a solo dev, it's not just anecdotal. It's the power structure underneath the infrastructure we use. Models can be banned, datacenters can go down, compute contracts can be terminated with 90 days' notice. Building on a single critical dependency remains a risk, whatever the layer.
And you, do you have a fallback strategy if your main model provider becomes inaccessible? Send me a message on Twitter/X or in the comments.
Alex
Key takeaways
GPT-5.6 Sol at $5 input vs Fable 5 at $10: the OpenAI/Anthropic price war is on
Gemini 3.5 Flash can control a browser, mobile app, or desktop, in production via the Google API
CLAUDE.md grows to 10 rules: loop engineering, separate verifier model, /goal command
SpaceX pulls $1.25B/month from Anthropic and $920M/month from Google for its Colossus datacenter
DeepSeek raises $7.4B at $50B valuation and builds its own infrastructure while the West regulates
