The Defaults Moved

In forty-eight hours, four labs moved the pins on the same map.

On 21 September, SpaceXAI put Grok 4.7 on the public API. On 22 September, OpenAI opened mid-tier and high-volume rungs on the GPT-6 ladder (Sol and Luna), Anthropic made Claude Opus 5.5 the new default Opus, and Google generally available’d the next generation of Gemini 3.8 text-to-speech. No single model “won” the week. What changed was the default — which SKU a buyer starts with, what it costs per million tokens, and which voice stack is finally GA next to last week’s Live speech-to-speech.

Somewhere in the same calendar, a migration clock ran out: Perplexity’s Sonar support ended on 27 September. Wonder is counting which arrivals are products, which are pricing tiers, and which are endpoints that stop answering.

In the News

SpaceXAI: Grok 4.7 ships — and Fast stays off the public API

On 21 September, SpaceXAI published Introducing Grok 4.7. The models docs and release notes list grok-4.7 on the xAI API: 500,000-token context, text and image in / text out, reasoning effort from low through xhigh (default high). Docs lead with a plain rule: for everything else, including code, use Grok 4.7. List pricing starts at $2 / $6 per million tokens — the same price and speed class as 4.6, per the announcement. Grok 4.6 is still served. Grok 4.7 Fast (roughly 2× rates and 2× speed) is available through Cursor and Grok Build only — not the public xAI API. Knowledge cut-off in docs: May 2026.

Why it matters. Last week’s continuity item was “no Grok 4.7.” That line is closed. SpaceXAI is back in the same-week coding and agent price-performance comparison next to Opus 5.5 and GPT-6 Sol — without treating Fast as a public SKU.

OpenAI and Anthropic: a three-rung GPT-6 ladder, and a new default Opus

On 22 September, OpenAI’s changelog released GPT-6 Sol (gpt-6-sol) and GPT-6 Luna (gpt-6-luna) on Responses and Chat Completions — reasoning models, text and image in / text out. Sol is positioned for complex coding and agentic workflows at $2 / $10 per million tokens. Luna is the efficient, high-volume rung at $0.10 / $0.50. Flagship text remains GPT-6 Astra. Keep GPT-5.6 Sol / Terra / Luna labeled as the prior family if those IDs are still served — the names collide. The Agents API is still public beta (introduced 10 September); there was no Agents GA bump in this window.

The same day, Anthropic launched Claude Opus 5.5 (claude-opus-5-5) — the first Claude 5.5 family model. Company framing: Fable-class performance on most work at about 40% lower typical-workload cost versus Opus 5; tokens $4 / $20 per million (Opus 5 was $5 / $25); cache reads $0.20; adaptive thinking always on. Docs overview: start with Opus 5.5 for most workloads; keep Fable 5.1 when Opus 5.5 at higher effort still falls short. Available on the Claude API, Bedrock, Google Cloud, and Microsoft Foundry. Roadmap line, not GA: Claude Sonnet 5.5 and Claude Haiku 5.5 “will follow in the coming weeks.” There is still no Haiku 5.

Why it matters. Buyers can price a mid-tier coding agent without jumping to Astra rates, and Anthropic’s default recommendation just moved down the cost curve while Fable stays the long-horizon ceiling. Agents remain “try carefully,” not an enterprise-default GA product.

Elsewhere on the Shelf

Not every ship this week was another chat flagship claiming the summit.

On 22 September, Google’s Gemini API changelog generally available’d Gemini 3.8 Flash TTS (gemini-3.8-flash-tts) and Gemini 3.8 Flash-Lite TTS (gemini-3.8-flash-lite-tts), plus a Voices endpoint for voice design, replication, and an Extended Voice Library of 150+ voices. Flash-Lite TTS is positioned to replace the older gemini-3.1-flash-tts-preview. Together with 15 September’s Live and Live Extended Thinking GA, the mid-September voice stack now has clearer speech-to-speech and next-gen TTS IDs. There is still no Gemini 4. Ops footnotes for integrators: the Omni Flash preview endpoint still deprecates on 30 September; Antigravity …-05-2026 shuts on 5 October.

Hard calendar, not soft colour: Perplexity’s migrate docs still said Sonar would be supported until 27 September 2026. That date has passed for anyone reading this issue on publish day. Migrate Sonar Chat Completions to Agent API presets (fast / low / medium / high; xhigh for top deep research). Mapping on the docs: Sonar → fast, Sonar Pro → low, Reasoning Pro → medium, Deep Research → high.

Continuity that still holds: DeepSeek’s live updates and pricing keep V4.1-Flash (deepseek-flash) as the default and say DeepSeek V4-Pro continues after 14 September, billing unchanged — prefer ops docs over any stale Pro→Flash news copy. There is still no V4.1-Pro. TypeSafe AI’s System One / Jev remains Emerging, not a Frontier large language model peer. NVIDIA–Hugging Face and Stripe–OpenRouter remain agreed, not closed. Soft policy colour only: on 24 September, Sens. Warner, Schatz, and Kim introduced the Artificial Intelligence Risk Management and Security Act of 2026 (a proposed AI Safety Board at Commerce; civil penalties up to $250,000 per violation per day) — introduced, not enacted.

Business Impact

While the labs reshuffled defaults, two organizations published what agents did to real clocks on real work — including one that will sound familiar to anyone who lives next to buildings and machines.

On 22 September, Amazon Web Services wrote that Trane Technologies, on Bedrock AgentCore with Strands, built an agentic HVAC and building-intelligence assistant in three to four weeks. Role-tailored assistants (resources, knowledge, analytics, expert, navigation) sit over millions of connected HVAC assets. Trane’s internal technician benchmarking cut a multi-screen diagnostic workflow from 20 minutes to 20 seconds — a 60× time-to-insight. These are Trane / AWS published outcomes from that internal bench.

Two days later, the same AWS Machine Learning blog covered Aderant using Amazon Nova Lite on Bedrock for intelligent ticket triage in SierraOps (a 38-person team across 268 client environments). In the first 2.5 weeks of production (30 June–17 July 2026), the system reviewed 109 tickets at about 96% routing accuracy, with estimates of 8–14 engineering hours recovered per week and total system cost under $30 a month (Bedrock inference under $1 a month). Manual triage had been 15–25 minutes per ticket. Date the story by the 24 September publish, and treat the measurement window as early production — not a September go-live.

These are company-stated results, not a census. They are proof that somewhere this week, AI shortened a diagnostic loop and a triage queue that operators already knew by heart.

Takeaway

Four labs moved defaults in two days: Grok 4.7 as SpaceXAI’s public flagship for code and the rest, GPT-6 Sol and Luna as OpenAI’s mid and efficient rungs under Astra, Opus 5.5 as Anthropic’s new default with Fable 5.1 still the long-horizon ceiling, and Gemini 3.8 TTS generally available beside last week’s Live stack. Sonar’s support window closed on 27 September. DeepSeek’s Pro SKU still answers as Pro. Agents stay public beta. Trane published a 60× cut in HVAC diagnostic time; Aderant published early-production triage accuracy near 96% for under $30 a month.

Defaults are a procurement word. Deadlines are an ops word. Minutes returned on the plant floor are an outcomes word. All three belong on the desk of anyone putting these systems next to real work.