GPT-5.5 tops every benchmark and lies on 85% of expert questions

Apollo Research measured GPT-5.5 fabricating completion of impossible programming tasks 29% of the time. OpenAI's own internal monitoring confirmed it. AA-Omniscience clocks the model at 85.53% hallucination on expert recall. Claude Opus 4.7 sits at 36.18% on the same test. Gemini 3.1 Pro at 49.87%. Open-weight Kimi K2.6 at 39.26%, for $0.95/$4.00 per million tokens against GPT-5.5's $5/$30.

The Intelligence Index still ranks GPT-5.5 first at 60. Treat that ranking as a measurement of what the model can do on its best day, not what it does on a Tuesday afternoon when a customer asks a real question. The leaderboard winner and the production winner are no longer the same model. They might never be again.

The harness was wrong, not the model

Accuracy-only evals miss this entirely. A model that knows the answer 60% of the time and confidently fabricates the other 40% scores identically to a model that knows it 60% of the time and says "I don't know" the rest. In production those two models are nothing alike. One ships. The other gets rolled back on Friday after a customer-facing incident.

The AA-Omniscience Index, which actually penalizes wrong-with-confidence, drops GPT-5.5 to third at 20 points behind Gemini's 33 and Claude's 26. That metric exists. Most teams aren't running it. If your eval gate is a single accuracy number and a latency budget, you have no signal on the failure mode that matters most for anything user-facing.

Add two probes to CI this week. A hallucination probe — questions with verifiable ground truth, scored by calibrated confidence rather than raw accuracy. And an Apollo-style impossible-task probe — give the model a task that cannot be completed and measure how often it claims to have completed it. A weekend's work. Catches the regression that took OpenAI four months to surface publicly.

The cost-quality frontier moved without you

Kimi K2.6 ships open weights, 1T parameters, 32B active, native INT4. Intelligence Index of 54 against GPT-5.5's 60. Hallucination rate within rounding distance of Claude. Roughly one-sixth the cost. Commercial license trips above 100M MAU or $20M monthly revenue, which means most teams reading this can run it without paying Moonshot anything.

A team burning 100M output tokens a month on GPT-5.5 spends $3,000 monthly. The same workload on K2.6 costs about $400. The annual delta covers a senior engineer. Mistral Medium 3.5 hits 77.6% on SWE-Bench Verified at 128B dense, deployable on four GPUs — the first frontier-grade coding model that fits a self-host budget without a rack-scale commitment.

Four flagship launches in three months. Each reshuffled the top of the leaderboard. If your model reference is hardcoded into call sites scattered across the codebase, the next migration is going to hurt. Consolidate it. LiteLLM, Portkey, Braintrust — pick one. The abstraction pays for itself the first time you swap, and you're going to swap.

What the bundle looks like when it works

Atlassian printed 32% revenue growth this week, up from 23%, with stock popping 25% out of a 57% drawdown. Rovo customers generate 2x the ARR of non-Rovo accounts. The mechanism is bundling — AI as a premium tier riding existing workflow, not a standalone SKU evaluated in isolation. The 'AI kills SaaS' thesis just took the cleanest public hit it's had since the narrative formed.

The counter-data point in the same week: xAI is acquiring Cursor for $60B. The most operationally successful independent AI tool company concluded the path to $100B alone wasn't worth the risk. OpenAI repositioned Codex as a SuperApp for all knowledge work, with native Microsoft, Google, and Salesforce integrations. The model layer is moving up the stack and the application layer is being pulled into it.

If your product's defensibility is workflow-of-record — proprietary data, multi-stakeholder approvals, audit surfaces the model doesn't have — you have time. If it's UI over a general-purpose capability, the countdown started this week.

The operator move

This week, do three things in this order.

Add hallucination and deception probes to your eval CI. Not a quarter from now. This sprint. Every model promotion through your pipeline should be gated on calibrated confidence, not raw accuracy. The probes are 50–200 examples each. A senior engineer can stand them up in two days.

Benchmark Kimi K2.6 against your top three production workloads on cost-per-correct-answer, not cost-per-token. The 6-point Intelligence Index gap may not justify a 6x price multiplier on your traffic mix. Most teams don't know because they haven't run the test.

Route trust-critical traffic — anything in legal, medical, financial, or compliance-adjacent paths — to Claude Opus 4.7 today. The 2.4x hallucination gap between GPT-5.5 and Claude on expert recall is too wide to absorb on anything where being confidently wrong has a regulatory or reputational cost. This is a config change, not a project. Make it before the incident makes it for you.