Your Analytics Copilot Survived Opus 4.8. The Claude 5 Upgrade Is Where It Breaks
The Monday after a model upgrade is quiet. Your weekly revenue report runs through Claude Code the way it always has. It queries the warehouse, writes the summary, lands in the inbox on time. Nothing errored, so nothing gets checked. That quiet is the trap. A model-family upgrade is a data migration, not an app update. The most important component in the pipeline that produces your numbers just changed, and the checks you built around the old component did not move with it. What breaks is not the model. It is your calibration: the prompts, thresholds and double-check habits you tuned, over months, to the previous model's specific weaknesses.
Seven weeks, three models, two deadlines
Anthropic shipped the Claude 5 family in three releases over seven weeks. Claude Fable 5 arrived on June 9 at $10 per million input tokens and $50 per million output tokens, in a new Mythos class that Anthropic places above Opus. Claude Sonnet 5 followed on June 30 at an introductory $2 and $10. Claude Opus 5 landed on July 24 at $5 and $25, the same sticker as Opus 4.8.
If you run analytics through Claude Code, two other dates matter more than the launch days. Opus 4.1 was retired on August 5. Anything still pinned to it now fails outright, which at least announces itself. And Sonnet 5's introductory pricing ends on August 31, when the standard $3 and $15 takes over. A volume workload you moved onto the cheap tier in July costs 50 percent more from September. The launches got the coverage. The deadlines are what reach an invoice.
View data table
| Date | Lane | Event |
|---|---|---|
| Jun 9, 2026 | Launch | Claude Fable 5 ships at $10 input / $50 output per million tokens, in the new Mythos class above Opus |
| Jun 30, 2026 | Launch | Claude Sonnet 5 ships at $2/$10 introductory pricing |
| Jul 24, 2026 | Launch | Claude Opus 5 ships at $5/$25, the same price as Opus 4.8 |
| Aug 5, 2026 | Deadline | Claude Opus 4.1 retired; requests to it fail |
| Aug 31, 2026 | Deadline | Sonnet 5 introductory pricing ends; standard $3/$15 applies from September 1 |
The report still runs. That is the wrong test
The dangerous changes in this family are the ones that never throw an error. Opus 5 thinks by default. A request that never set the thinking parameter now spends reasoning tokens before it answers. The output cap covers reasoning plus answer together, so a tightly sized report job can come back truncated mid-table. Sonnet 5 uses a new tokenizer, and Anthropic's migration guide puts the shift at roughly 1.0 to 1.35 times as many tokens for the same text. Per-token prices did not move. Everything you measure in tokens did: budgets, context limits, the cost dashboard you reconciled last quarter.
There is more in the same category. Opus 5 draws on its own rate limits, separate from the shared Opus 4.x pool, so moving traffic neither frees old headroom nor inherits it. Prompt caches are scoped per model, so the first day on a new model runs cold and costs more. Every one of these changes is documented. None of them stops a report from completing. For an analytics workload, that is exactly the failure profile to fear: the deliverable arrives on schedule, and the machine underneath it has different economics and different habits.
What actually breaks is your calibration
Calibration is everything you added around the model to compensate for the model: the double-check instruction, the severity threshold, the re-run rule. It accumulates one fix at a time, and nobody writes it down as a system. Then the model changes, and the fixes invert.
Anthropic's own migration guide is unusually blunt about this. It tells Opus 5 users to delete their verification scaffolding, because the model now verifies its work unprompted. Kept in place, those instructions produce over-verification: slower runs and higher spend for no added accuracy. The same guide warns that conservative reporting filters now bind literally. Tell the model to flag only material variances, and it will investigate as thoroughly as before, then withhold findings it judges below your stated bar. Fewer flags in the weekly anomaly report can look like cleaner data when it is actually lost recall. Nothing fails. Your safety net just stopped netting, on the same day the model got smarter.
Run it like the data migration it is
View data table
| Step | What happens |
|---|---|
| 1. Inventory | Find every place a model ID or alias lives: settings files, agent definitions, scheduled jobs, API scripts |
| 2. Baseline | Run 10 to 15 golden questions with settled answers on the old and new model; diff the numbers, not the prose |
| 3. Remap | Workhorse tasks to Opus 5 at the old price; volume to Sonnet 5; Fable 5 only where errors cost more than tokens |
| 4. Promote | Move read-only reports first, hold board-facing numbers until baselines match, then delete the old calibration |
Step one, inventory the pins. A model ID hides in more places than the settings screen: global settings files, per-agent definitions, scheduled jobs, the scripts that call the API directly. I ran this inventory on my own operation. The server that writes and publishes the essays on this site fires six scheduled Claude Code jobs: this essay loop, a daily operations brief, a weekly data-quality run, and weekly memory, security and LinkedIn-drafting jobs. The search for pinned model IDs came back empty. Every job inherits one line in one settings file, which means a single edit re-based six unattended jobs at once. That is the quieter of the two failure shapes. Scattered pins strand jobs on old models silently. One shared pin flips jobs you forgot you had. You do not know which shape you run until you look.
Step two, baseline before you believe. Collect ten to fifteen golden questions: queries whose answers are already settled, like last month's closed revenue or a cohort count the board has already seen. Run them on the old model and the new one, and diff the numbers, not the prose. The prose will differ. It always does. The numbers must not.
Step three, remap by tier, not by name. Opus 4.8 to Opus 5 is the easy call, a stronger model at an identical price. Sonnet 5 is the volume tier, if you re-baseline your token budgets for the new tokenizer before September. Fable 5 is not the default upgrade. At twice the price of Opus 5, it is for the questions where being wrong costs more than the tokens do.
View data table
| Model | Input | Output | Note |
|---|---|---|---|
| Claude Fable 5 | $10 | $50 | Mythos-class tier above Opus |
| Claude Opus 5 | $5 | $25 | Same price as Opus 4.8 |
| Claude Sonnet 5 | $3 | $15 | Introductory $2/$10 until Aug 31, 2026 |
| Claude Haiku 4.5 | $1 | $5 | Carried over from the previous family |
Step four, promote deliberately, then delete. The old models stay in service, so nothing forces a same-day cutover. Move the read-only reports first. Hold anything board-facing on the old model until its golden questions match. And once they match, finish the migration: strip out the old calibration, the double-check prompts and the workaround instructions. On this family they are not harmless leftovers. They are active misconfiguration.
Here is the position this essay should be graded on. The Claude 5 incidents that reach postmortems in analytics teams this year will be calibration failures, stale prompts and skipped baselines, not the new models computing worse answers than their predecessors on well-tested pipelines. If properly baselined pipelines degrade on this family by mid-2027, I read it wrong. Either way, the clock on your migration is not abstract. August 31 is three weeks out. Run the inventory before the pricing runs it for you.