The Second-Source Test: Pricing Your AI Stack Against a Live Alternative
The Departure
Two frontier models spent this week moving in opposite directions, and for once it wasn't theoretical. GPT-5.6 (Sol, Terra, and Luna) just finished its first full week genuinely public, with real prices and real verdicts attached. The same week, Anthropic moved the included Fable 5 deadline again: July 7 became July 12 became July 19. Third move in six weeks.
If you followed #023 through #026, none of this asks you to do anything. You already treat the model as borrowed. What's new is smaller and more useful: for the first time, you can price the alternative on real numbers instead of a launch promise.
The Co-Pilot
Tool: Claude Opus 4.8 via claude.ai. Still the stable workhorse. It hasn't moved once while every Fable 5 date has, so run the analysis here, not on a model whose access keeps shifting.
The Use Case: The second-source test. Paste in the AI work you actually run, and get it priced across four real options with one honest call: stay put, add a backup, or switch.
The Prompt: "You are a fractional CFO. Here's the AI work I actually run each week: [paste your tasks, how many times a week, and rough input/output length for each, e.g. 'drafting 5 client proposals a week, about 2,000 words in and 1,500 words out']. Price that exact monthly workload across four real options: (1) GPT-5.6 Terra at $2.50 per million input tokens and $15 per million output; (2) Claude Sonnet 5 at $2/$10 through August 31, then $3/$15; (3) Claude Opus 4.8 at $5/$25; (4) metered Fable 5 at $10/$50 as the ceiling. Then do three things: (a) name the single option that fits this workload best on cost; (b) tell me honestly whether switching is worth the setup and re-testing friction at the dollar difference involved, or whether the gap is too small to bother; (c) give me one call, stay put, add a second source as a backup, or switch outright, with the monthly dollar difference attached."
New here, or did this get forwarded to you? You don't need any prior dispatch. Paste roughly what you run each week and how often, and run it cold.
And it runs on your phone. Type it into the Claude app from a gate or a back seat, run it, and screenshot the call.
The Upgrade
Topic: The week in models, a moved deadline, a real launch, and a wildcard price
The deadline moved a third time. July 7 became July 12 became July 19 (per Ben's Bites and The Rundown this week). That's through 11:59:59 PM PT on the 19th, for Pro, Max, Team, and premium Enterprise seats. Two separate things share that date: you can use Fable 5 for up to 50% of your weekly subscription limit at no extra cost, and Claude Code's 50% higher weekly limits reset at the same time each day. After that, Fable moves to metered usage credits ($10/$50 per million tokens) unless the date moves again. To be fair to Anthropic, this reads as a compute-supply problem: they say they plan to restore included access once they have the capacity. But even Ben Tossell of Ben's Bites, who writes one of these newsletters, put it plainly: "This dangling-a-carrot move doesn't make me trust their deadlines anymore." Don't plan around July 19 either.
GPT-5.6's first week was real and messy. The prices are confirmed from OpenAI's own docs: Sol $5/$30 (same as GPT-5.5), Terra $2.50/$15, Luna $1/$6 per million tokens. On the verdicts, take these as early and hedged. An early tester via Ben's Bites called Sol "not smarter than Fable but so much more reliable." The Rundown puts Sol slightly below Fable 5 on Artificial Analysis's Intelligence Index but ahead on agentic coding. Ben's Bites' own first-week read is that Terra is basically a GPT-5.5 replacement with minor gains. The rollout itself was rough: OpenAI reset usage limits four or five times and temporarily lifted rate limits while it fixed bugs (per Ben's Bites). So it's real, public, and priced to compete. It's also week-one messy. Test it on something low-stakes before you move anything that matters.
One more, and one still open. The cheapest number on the table this week isn't from OpenAI or Anthropic. xAI's Grok 4.5 launched around $2/$6 per million tokens (per The Rundown), pitched as Opus-class. It's one week old, developer-leaning, and unproven for real business work, so watch it, don't move to it. And to keep two threads honest: there's still no word on whether Copilot Cowork's usage billing applies inside the new M365 bundles, and the reported Microsoft in-house-models push went quiet this week. Nothing to act on in either.
The Landing
Ten minutes, today. Open Claude Opus 4.8. Paste the AI work you actually run each week. Run the second-source test above and get your one call with the dollars attached.
Here's the throughline. #023 said to treat intelligence as borrowed. #026 put a dollar figure on the dependency. This week, the market finally handed you a real second source to price against. Don't wait until July 19 to learn what your stack costs without the safety net.
Safe travels,
Marcellus