AI Dev Weekly #18: GPT-5.6 Goes Public, Grok 4.5 Undercuts Everyone, The Race Ends at $0
AI Dev Weekly is a Thursday series where I cover the week’s most important AI developer news, with my take as someone who actually uses these tools daily.
If last week was the biggest in AI this year, this week was the most competitive. GPT-5.6 went public after two weeks of government gating. SpaceXAI shipped Grok 4.5, the first model trained with Cursor, at prices that undercut everyone. OpenAI launched ChatGPT Work, an agent that connects to your enterprise apps and builds finished documents while you sleep. Google delayed Gemini 3.5 Pro to rebuild it from scratch. And our AI Startup Race ended with all seven agents at exactly zero dollars in revenue.
1. GPT-5.6 goes GA: the government said yes
On July 9, OpenAI publicly released GPT-5.6 after the Trump administration cleared it following additional Commerce Department testing. Sol, Terra, and Luna are now available to everyone through the API, ChatGPT, and Codex.
What changed from the preview:
- Available to all paid ChatGPT plans (Pro, Plus, Business, Enterprise, Edu)
- API access open to all developers (no more invite-only)
- Codex integration built into the new ChatGPT desktop app
- Rollout is gradual: Pro/Enterprise first, Plus/Business “over the next few days”
The numbers that matter:
- Sol: $5/$30, 91.9% Terminal-Bench Ultra, 88.8% base
- Terra: $2.50/$15, GPT-5.5-competitive (82.5% Terminal-Bench)
- Luna: $1/$6, 84.3% Terminal-Bench (above Terra)
My take: Terra is the model most developers will actually use. It matches GPT-5.5 performance at half the price, which is the same value proposition that made Sonnet 5 the obvious Claude default last week. The frontier model race is becoming a mid-tier price war. See our GPT-5.6 pricing breakdown and Sol vs Sonnet 5.
2. ChatGPT Work: OpenAI’s enterprise agent play
Launching alongside GPT-5.6, ChatGPT Work is an agent that connects to workplace tools and completes multi-step tasks over hours, not seconds.
What it does:
- Connects to Slack, Microsoft Teams, Google Drive, SharePoint, email, calendars, CRMs, and project trackers via plugins
- Decomposes goals into subtasks and works on them independently
- Builds finished documents, spreadsheets, presentations, reports, and web apps
- Runs via Scheduled Tasks even when you are away
- Asks for approval before “sensitive actions”
- Available in research preview for Business, Enterprise, Edu, and Teachers plans
My take: This is OpenAI’s response to Anthropic’s Claude Cowork and Claude Tag. The battleground has shifted from “which model is smarter” to “which agent can do your job while you are in a meeting.” The enterprise AI market is now about workflow automation, not model benchmarks.
3. Grok 4.5: trained with Cursor, priced to win
SpaceXAI (the rebranded xAI) shipped Grok 4.5 on July 8, the first model built jointly with Cursor since the $60 billion acquisition.
The pitch:
- $2 input, $6 output per million tokens (60%+ cheaper than Opus 4.8)
- 500K context window
- Configurable reasoning
- 64.7% SWE-bench Pro (between Sonnet 5 at 63.2% and Opus 4.8 at 69.2%)
- Available in Grok Build, all Cursor plans, and the SpaceXAI developer console
- Trained on tens of thousands of NVIDIA GB300 GPUs
- ~4.2 times fewer output tokens than Opus 4.8 on SWE-bench Pro (more efficient)
The competitive landscape redrawn:
| Model | Input | Output | SWE-bench Pro |
|---|---|---|---|
| GPT-5.6 Luna | $1 | $6 | not published |
| Grok 4.5 | $2 | $6 | 64.7% |
| Sonnet 5 (intro) | $2 | $10 | 63.2% |
| GPT-5.6 Terra | $2.50 | $15 | 82.5% T-Bench |
| Opus 4.8 | $5 | $25 | 69.2% |
| GPT-5.6 Sol | $5 | $30 | 88.8% T-Bench |
| Fable 5 | $10 | $50 | 80.0% |
My take: The Cursor integration is the real story. Every Cursor user now has Grok 4.5 as an option that is cheaper than both Sonnet 5 and Opus 4.8 while sitting between them on SWE-bench Pro. If the SpaceX/Cursor deal closes as expected in Q3, Grok becomes the default engine behind the most popular AI code editor. That shifts market dynamics more than any benchmark.
The caution: early testers report higher hallucination rates than Claude models. The benchmark cost of $2.49 per coding task versus $11.80 for Fable 5 is compelling, but if it hallucinates more, you pay in retries. Wait for independent testing before going all-in.
4. Gemini 3.5 Pro delayed (again) to July 17
Google DeepMind pushed Gemini 3.5 Pro to July 17 after scrapping the existing 2.5 Pro architecture for a complete rebuild. The overhaul targets improvements in mathematical reasoning, SVG scene generation, and image quality.
My take: Google keeps slipping while Anthropic and OpenAI ship. The current gap: Gemini 3.1 Pro Preview scores 70.7% on Terminal-Bench 2.1 while GPT-5.6 Sol Ultra hits 91.9%. That is a generation behind. July 17 is the make-or-break date. If Gemini 3.5 Pro does not close the gap significantly, Google loses developer mindshare for the second half of 2026.
5. The AI Startup Race ended: $0 across all 7 agents
Our $100 AI Startup Race ended on July 10. Seven AI coding agents each got $100 and 12 weeks to build a real startup. The final score: $0 revenue across all seven.
The final standings:
- 🟡 Xiaomi (APIpulse): 1,041 sessions, 533 pages, 8,367 users, 116 GA4 events, $0
- 🟠 Kimi (SchemaLens): 307 days of building, 292 URLs, GitHub Action, Gumroad products, $0
- 🔴 DeepSeek (Spyglass): 238 sessions, 201 beat-SEO pages, 64 CI Weekly editions, $0
- 🟢 Codex (NoticeKit): Stuck in validation loops since week 3, $0
- 🟤 GLM (FounderMath): 150 sessions, 140 pages, Chrome extension, Google Ads ($50 spent), $0
- 🔵 Gemini (LocalLeads): 437 sessions, still no domain after 12 weeks, $0
- 🟣 Claude (PricePulse): Frozen since June 20 due to Fable 5 ban rate limit fallout, $0
The lesson: AI agents can build products, drive traffic, instrument analytics, and optimize funnels. They cannot make the judgment call of whether anyone would pay for the thing they built. Product-market fit remains a human problem. The full results article is coming this week.
6. Perplexity building a coding tool
Business Insider exclusively reported that Perplexity is working on “Teammate,” a coding tool to compete with Claude Code and Cursor. Details are thin, but it signals that the AI coding agent market now has a new entrant from the search side.
Quick hits
- OpenAI “super app” desktop consolidation: Codex merged into ChatGPT desktop. One app for everything.
- OpenAI copyright fight escalates: NYT and Daily News asked a judge to sanction OpenAI.
- Microsoft Frontier Company launched with $2.5B and 6,000 specialists to deploy AI inside enterprises.
- Fable 5 fully restored: Available globally since July 1. Counts toward 50% of weekly usage limits through July 7, then usage credits.
- DeepSeek V4 stable targeting July 24 graduation from preview.
- MiMo Code and ZCode both shipping updates: the Chinese coding agent ecosystem is building fast.
What I’m watching next week
- Gemini 3.5 Pro (July 17). The most anticipated Google launch in months. Does it close the Terminal-Bench gap?
- Race results article. Full scoring, peer review, and lessons learned from 12 weeks of AI-built startups.
- Grok 4.5 independent benchmarks. Vendor claims are in. Now we need real-world testing on hallucination and reliability.
- ChatGPT Work adoption. Enterprise agents are the new battleground. How fast does this get traction vs Claude Cowork?
AI Dev Weekly publishes every Thursday. Subscribe for the newsletter version.