Two frontier stacks hit the public API within 24 hours. OpenAI finally ships GPT-5.6 Sol (plus Terra and Luna) after a US government delay. SpaceXAI drops Grok 4.5 co-trained with Cursor. Microsoft, meanwhile, starts routing Excel and Outlook Copilot traffic off OpenAI and Anthropic onto its own MAI models.

If you build coding agents, today is not a newsletter day. It is a model-swap day.

GPT-5.6 Sol, Terra, Luna — public after the gate

OpenAI delayed the broad GPT-5.6 rollout in June after Washington asked for early access under the frontier-model oversight framework. On July 7 the company said the green light was in: Sol, Terra, and Luna launch publicly on Thursday, July 9, with global preview access expanding first.

Primary signal from OpenAI on X: GPT-5.6 Sol is the flagship; Terra and Luna are the cheaper tiers. Sam Altman backed it with a short “happy building” note. Coverage from The Straits Times and Reuters frames this as the same national-security review cycle that briefly yanked Anthropic’s Fable/Mythos and then put them back under tighter classifiers.

What matters for agent builders:

• Sol — hardest problems: long-horizon coding, security research, science. Early chatter mentions an Ultra subagent mode that spins parallel workers for one prompt. Treat that as “swarm reasoning in the API” until the system card is public.
• Terra — everyday workhorse; marketed as GPT-5.5-class capability at roughly half the cost.
• Luna — high-volume, low-latency tier for drafting and cheap agent loops.

The government gate is the real story under the marketing. Frontier releases are no longer “we ship when the evals pass.” They are “we ship when Commerce is done looking.” Same pattern we covered when Fable 5 returned after export controls.

Grok 4.5: Cursor-trained, cheap, fast

On July 8, SpaceXAI published Introducing Grok 4.5. This is the first joint model with Cursor after the SpaceXAI–Cursor combination, and the first Grok release since SpaceXAI went public.

Hard numbers from the official post:

Benchmark chart on the announcement page puts Grok 4.5 mid-pack to strong on engineering evals (DeepSWE, Terminal Bench 2.1, SWE Marathon, SWE Bench Pro) — not always first against Fable max or GPT-5.5 xhigh, but consistently competitive while costing a fraction of Opus-class list prices. Axios quotes Musk calling it “Opus-class… but faster, more token-efficient and lower cost.”

The agent-relevant angle is training data, not leaderboard cosplay. SpaceXAI says the model was co-trained with Cursor on multi-step software engineering rollouts — hours-long agent trajectories, automated grading, asynchronous RL across tens of thousands of GB300s. That is the same distribution your coding agent actually samples: tools, patches, failed tests, retries.

# Grok 4.5 via SpaceXAI API (from official docs)
curl -s https://api.x.ai/v1/responses \
  -H "Authorization: Bearer $XAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "grok-4.5",
    "input": "Find and fix the bug, then explain it: function median(a){a.sort();return a[a.length/2]}"
  }'

Limited free usage is live in Grok Build and Cursor. If you already pay for Cursor seats, you just got a new default without a procurement ticket.

Microsoft starts eating its own OpenAI bill

Separate but related: Bloomberg reported (July 7) that Microsoft is routing tens of thousands of weekly Copilot prompts in Excel and Outlook through in-house MAI models instead of OpenAI and Anthropic. Summaries at MLQ and Yahoo Finance rehash the same core fact: this is a cost and dependency move, not a product keynote.

For agent platforms sitting on Microsoft 365, the implication is boring and important: model identity behind “Copilot” is no longer a single vendor. Routing will keep shifting toward MAI for high-volume, lower-stakes Office tasks while frontier OpenAI/Anthropic calls stay for harder work. If your product assumes “Excel AI = GPT,” update the assumption.

Stack notes for today

If you run coding agents:

1. Try Grok 4.5 as the cheap parallel worker. At $2/$6 and high TPS, it is a natural subagent for multi-file refactors, test generation, and office-adjacent automation while you keep a Sol/Fable-class model for planning and merge judgment.
2. Re-price Sol vs Terra. Once Sol/Terra/Luna list prices settle, rebuild your router. Terra-as-default + Sol-on-escalation is the obvious pattern OpenAI is selling.
3. Watch the government gate. Fable returned with stricter classifiers. Sol launches after Commerce review. Policy latency is now part of model risk — budget for delayed rollouts on the next frontier drop.
4. Do not assume Microsoft routes are stable. MAI in Excel/Outlook is the canary; more Office surfaces will follow.

Hardware side note if you are still local-first: denser open models (Qwen 3.6 27B class) keep showing up in agent harnesses for offline loops, but today’s news is API economics, not another GGUF release.

Useful gear if you are still wiring a dual-GPU agent box: a used NVIDIA A6000/6000 Ada for local runners, a solid USB-C dock for laptop harnesses, and a mechanical keyboard that survives all-night agent runs. For team seats, Cursor remains the path of least resistance to Grok 4.5; Claude Code / Codex stacks still map cleanly onto Sol when API access unlocks.

Takeaway

July 9 is the day the gated GPT-5.6 family finally hits public rails and SpaceXAI proves the Cursor acquisition was a training-data play, not just a product bolt-on. Microsoft’s MAI routing is the quiet third story: hyperscalers will keep eating their own inference bill wherever quality is “good enough.”

Router configs matter more than brand loyalty this week. Sol for the hard path. Grok 4.5 for the cheap, fast swarm. MAI for Office bulk. Everything else is marketing.

Some links are affiliate links (Amazon tag spaghettistor-20). We may earn a commission at no extra cost to you.