The Real Cost of Depending on One AI Model
The Departure
Last week, I called July 7 the "hard" cutoff for Fable 5 access to be included. It didn't survive the week.
Anthropic quietly pushed it to July 12 (per Anthropic, reported this week by The Rundown and Superhuman AI). That's the second time in five weeks a stated Fable 5 deadline has moved. First, the government pulled the model before "free through June 22" ever closed. Now the vendor moved its own line.
If you followed #023 and #025, kept the artifact, and ran your real work on Opus 4.8, then none of this applies to you. That's the whole point. So this week I'm putting real dollars on what, depending on a model like that, actually costs.
The Co-Pilot
Tool: Claude Opus 4.8 via claude.ai. The stable workhorse through this whole mess. It hasn't moved once, while every Fable 5 date has.
The Use Case: A paste-and-go dependency-cost calculator. Tell it what you actually use Fable 5 for and how often, and you get back what that workload costs on metered Fable 5, Opus 4.8, and Sonnet 5, plus one honest call.
The Prompt: "You are a fractional CFO. Here's how I've been using Claude Fable 5: [paste the tasks, roughly how many times a week, and rough input/output length, e.g. 'drafting 5 client proposals a week, about 2,000 words in and 1,500 words out each']. If my subscription-included access ends, Fable 5's metered rate is $10 per million input tokens and $50 per million output tokens. My subscription already includes Claude Opus 4.8 ($5 in / $25 out per million) and Claude Sonnet 5 ($2 in / $10 out through August 31, then $ 3 in/$15 out). Do this: (1) estimate my monthly cost if I kept this exact workload on metered Fable 5; (2) estimate the same workload on Opus 4.8 and on Sonnet 5; (3) tell me honestly whether Fable 5's extra capability is worth the premium for what I actually use it for, or whether Opus or Sonnet gets me most of the result for a fraction of the cost; (4) give me one call: keep using Fable 5 through whatever window stays open, switch now, or admit I never depended on it enough for this to matter."
New here, or did this get forwarded to you? You don't need any prior dispatch. Paste roughly what you use Fable 5 for 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. One more if you're on a Max plan: Anthropic's own Cowork feature (its own tool, not Microsoft's similarly named Copilot Cowork) just landed on web and mobile, so this whole exercise now runs the same wherever you are.
The Upgrade
Topic: Two frontier models moved in opposite directions this week
The deadline moved again. July 7 was supposed to be the last day of included Fable 5 access. Yesterday, that became July 12, quietly, with no real explanation (per Anthropic, as reported by The Rundown and Superhuman). The second moved deadline in five weeks. It's unclear whether the same 50%-of-weekly-limits throttle still applies during the extension, so don't build anything load-bearing on it. And don't treat July 12 as more final than July 7 was. Build for the pattern, not the date.
The model that was too dangerous for the public is now yours, and it's cheaper. GPT-5.6's Sol, Terra, and Luna tiers go public today after the Commerce Department cleared the review (per Engadget, MacObserver, and Neowin). Sol was locked to about 20 government partners three weeks ago. Its preview rate is $5 in, $30 out per million tokens: half of Fable 5's metered $10/$50. Terra promises GPT-5.5-level output at half the cost. So in one week, one frontier model became expensive and metered, while another became cheap and public. That's the sharpest argument yet for keeping your workflows model-agnostic. The prompt, the context file, and the artifact carry over. The model underneath is a price-and-availability call you re-make each quarter, not a marriage. One honest caution: launch dates have slipped all month (twice for Fable, once for GPT-5.6 itself), so confirm it actually landed in your account before you build on it.
Watch this: the model behind your Copilot bundle may be changing too. Reports this week say Microsoft's Mustafa Suleyman is pushing to move Copilot's underlying models onto Microsoft's in-house "MAI" models, to "ultimately eliminate" its Anthropic bill (per The Rundown). Reportedly, and directional. Nothing's confirmed as changed yet, and it doesn't resolve the Cowork billing question from #025 (still open after another look at Microsoft's pages). But it's a reason to keep watching what's actually running under the hood of the bundle you may have just chosen.
Skipping this week: Meta's Muse image and video models and Sam Altman's sovereign-wealth-fund proposal. Both are interesting, but neither changes what you do this week.
The Landing
Ten minutes today. Open Claude Opus 4.8. Note what you've actually used Fable 5 for this week and how often. Run the dependency-cost calculator above. Get your keep, switch, or "never depended on it enough to matter" call with the real-dollar comparison attached.
Do it before the next deadline moves on you. At this point, one will.
Safe travels,
Marcellus