Between June and October 2026, Anthropic, OpenAI and Google each released a new flagship or price tier. If you buy AI conversations for your business, the question is not which model leads this month. It is what to do when the lead changes every quarter.
What the numbers say
Capability gets cheaper fast. Epoch AI estimates the cost of reaching a fixed level of capability has fallen about 13 times a year since 2023. One example: the same 75% score on a hard science benchmark cost about $0.30 a question in January 2025 and about $0.0004 in 2026.
But the price of a named model can rise. Google has scheduled price doubles for Gemini 4 Argon (1 December 2026) and its Flash models (1 January 2027).
Rankings disagree. Artificial Analysis ranks Claude Opus 5.5 first this month. Epoch AI ranks GPT-6 Astra first. The answer depends on the test.
Models retire. Anthropic told developers that Claude Sonnet 4.5 retires on 30 November 2026, with about two months' notice.
A flagship can disappear. In June 2026 one leading lab's flagship models were suspended for all users for 19 days under a US export-control directive.
Open models are close behind. Epoch puts open-weight models about four months behind the closed frontier on general capability.
The stay-flexible checklist
Keep prompts and knowledge separate from the model. Never hard-wire one model name.
Test on your own conversations. Build a set of 100 or more real calls and chats, including hard and angry ones, and re-run it on every new model.
Set guardrails: a minimum quality score, a maximum cost per resolved conversation and a maximum delay before the first reply on voice.
Route by task, with a fallback. A small model for simple requests, a larger one for hard ones, and an automatic switch to a second provider on errors.
Contract for the exit: price protection, 90 days or more of notice of model changes, data portability and no commitment to one model.
Rehearse a model swap. Watch deprecation and price pages monthly and practise a swap before you are forced to make one.
Why it matters for voice
Voice is a latency problem first. Independent tests show roughly 0.6 to 1.4 seconds to first audio for the leading real-time voice models, and the fastest settings can cost accuracy. Judge on real call recordings, not on a headline figure.
Read the full report
Our industry report The AI model race: pick outcomes, not favourites has the release timeline, the myths and every source.
View the full report on our Resources page →
This article summarises published research and vendor announcements as of 23 October 2026 and is for general information only. Model names, dates and prices change quickly. Benchmark results depend on the test and settings used.
