Your Model Spreadsheet Needs Two New Columns

The spreadsheet everyone kept had two columns. That’s the problem.

Your model doc probably looks like every other one I’ve seen for the last eighteen months. Provider on the left. Input price, output price, a couple of benchmark scores across the top. Maybe you threw in a context-window column. You pick the cheapest box that clears your workload and ship.

That spreadsheet just got two new columns whether you like it or not.

June handed you nineteen days of frontier downtime and a wave of real-time safety classifiers quietly downgrading queries. Meet the new columns: access continuity risk and safeguard false-positive rate. If you don’t score them, you’ll learn about them the hard way — in an incident call with a customer waiting.

I’ve written about switching costs and the three-layer routing pattern I run in prod. This is the payoff. The point isn’t “drop frontier models.” The point is: don’t leave a core layer of your stack at one provider’s discretion.

What actually broke, in one paragraph

Anthropic’s Fable 5 and Mythos 5 were dark for nineteen days under a government export-control order. In that same window, OpenAI ran GPT-5.6 Sol as a government-coordinated preview with roughly twenty vetted orgs before GA. Anthropic also tiered the same underlying model as Fable (open) and Mythos (approval-gated). 120+ security researchers signed the freefable.org letter arguing defenders lost access to tools attackers still found ways around.

That’s the event. What matters is what it proved about the structure of this market.

New column one: access continuity risk

Your uptime SLA doesn’t cover this: a government order, a policy flip, a vetting hurdle your counsel can’t clear in ninety days, or a mid-quarter change to your tier’s rate limits.

Access to the frontier used to be gated by capital. Now it’s gated by trust credentials — NDAs, vetting, government coordination, partner-of-record status. If you build in a regulated vertical, you feel this already. A biosecurity founder friend has a Mythos-tier request sitting for six weeks. Their competitor — a Big Four — got approved in nine days.

Save the moralizing. Vendors have reasons to gate. Your job is simpler: the top of the model market isn’t merit-only anymore. Access is a variable. Model it.

Make it concrete. Ask your team: if our primary provider stops serving us tomorrow — for any reason — how many days do we keep the product up before revenue takes a hit? Nineteen days is now the empirical baseline for a shutdown. If your answer is “we’re down in an hour,” you don’t have a model strategy. You have a dependency.

New column two: safeguard false-positive rate

Fable 5 ships with real-time safety classifiers watching for cyber, bio, chem, and a few other categories. Flagged sessions get silently routed to a smaller, older model. Anthropic says it’s under five percent of sessions. That average sounds fine. Averages hide domains.

A SANS instructor went public saying normal defensive security work — analyzing malware behavior, writing detection rules — was getting flagged and downgraded. On Hacker News, a consumer reported getting blocked for asking about yellow-background cells in Excel because “yellow” + “background” + formatting language tripped a classifier. Different stakes, same mechanism your API traffic hits.

Here’s the nasty, specific paradox: the more your vertical’s vocabulary overlaps a flagged category, the higher your silent quality degradation. Security tools use attacker language. Medical tools use drug and dosage language. Legal tools discuss violence and fraud in the abstract daily. Fraud detection reads like a fraud manual by design. In these domains, your effective model quality is not the benchmark page. It’s the benchmark score times (1 – your domain’s flag rate).

Nobody publishes that number. You have to measure it.

The measurement most teams skip

Grab your top one hundred real production prompts (redact if you must). Run them on your current model. Capture the outputs. Look for downgrade tells: sudden refusals, new hedging, drops in structured-output compliance, or answers that feel one generation dumber than last month.

Now run the same hundred against the downgrade target the classifier would send you to. Compare quality. Weight the gap by your observed flag rate. That’s your real cost from the safeguard system. In security tooling I’ve measured, it’s non-trivial — high single digits of user-visible quality loss on flag-adjacent prompts.

Same principle as vendor benchmarks: their average isn’t your average. Build a golden set. Extend it for safety behavior.

What routing actually looks like now

I run three layers in prod. Cheap model for classification and extraction — Haiku, Luna, mini tiers. Frontier model for the 10–20% of requests that truly need depth. Fallback under both.

That fallback used to be insurance against deprecation and rate limits. Now it’s insurance against policy action too. That promotion means it has to live at a different provider, not just a different SKU from the same vendor. If Anthropic goes dark, Anthropic-Haiku goes dark with it.

The pricing math finally cooperates. Sol at $5/$30 per million tokens is exactly half of Fable’s $10/$50. Mini tiers — Luna at $1/$6, GPT‑4.1‑mini and Haiku in the same band — are cheap enough that a classifier layer costs almost nothing. Prompt caching takes another ~90% off repeated context. There’s no economic reason left to send classification traffic to a frontier model. The only remaining reason is architectural laziness.

If more than 80% of your traffic doesn’t need the frontier — and most teams overestimate that by 3–5x — you’re already past breakeven on building routing.

The June shutdown was a fire drill nobody scheduled

Here’s the reframe. The nineteen-day Fable outage wasn’t a black swan. It was a fire drill.

Anyone with an Anthropic dependency got forced through a migration: rewrite prompts for a new model family, tweak tool schemas, re-run evals, pause the roadmap. Teams with provider-abstracted routing barely flinched. Teams without it lost real days.

The lesson isn’t “Anthropic is unreliable.” The lesson is the category now includes forced migrations. Regulators will act again. Vendors will change terms again. Safety policies will tighten and loosen again. If your architecture treats any single provider as permanent infrastructure, you’re one policy memo away from a bad quarter.

Five things to run this month

Not a listicle. Homework — in order.

One. Run the fire drill. Assume your primary provider stops serving traffic tonight. Write down, hour by hour, what breaks and when. If you can’t answer, that’s project one.

Two. Build the domain false-positive test. One hundred real prompts. Primary model versus downgrade target. Measure the gap. Re-run quarterly — safety classifiers change silently.

Three. Extend your golden eval set with a flag-adjacent cohort. Vendor benchmarks won’t warn you about your vocabulary.

Four. Measure the real share of requests that need frontier reasoning. Most teams are off by 3–5x. That honest number is your routing ROI ceiling.

Five. Read your provider contracts for data retention and redistribution. Thirty-day retention is common. In some verticals, that clause alone is a compliance blocker you already signed.

The frontier is still worth using. It’s just not worth depending on. Add two columns. Spend one honest afternoon on your routing. Build for the market we have now — not the one from eighteen months ago when price and score were the only cells that mattered.