Both Labs Cut Prices. Your Bill May Not Drop.

Anthropic and OpenAI both shipped cheaper models today, hours apart. Anthropic says Claude Opus 5.5 matches its Fable 5.1 flagship at around 40% less to run than Opus 5. OpenAI says GPT-6 Sol and Luna come in 50% below the promotion currently running on GPT-5.6. Both companies say they are passing efficiency savings to customers. Whether that reaches your invoice depends on three things neither announcement leads with.

What was actually announced

Anthropic's Opus 5.5 is described as its most efficient model, performing at the level of Fable 5.1 while costing roughly 40% less to run than Opus 5, which launched in July. It is the first of a family, with Sonnet 5.5 and Haiku 5.5 expected over the coming weeks. Anthropic says it is passing the savings on through price cuts and rate limit increases.

OpenAI's GPT-6 Sol and Luna are offshoots of the GPT-6 Astra flagship released earlier this month, positioned as versions of Astra for everyday work. OpenAI attributes the reductions to improvements in caching and inference. Fortune reported that OpenAI's pricing comes out below Anthropic's, based on each company's own announcement rather than on independent testing.

Two flagship releases on one day, both cheaper, is genuinely good news for buyers. The structure underneath is where the money actually moves.

Three things that decide your invoice

The baseline. OpenAI's 50% is against a promotion currently running on GPT-5.6, and published rates list that Sol promotion as available at least through November 21, 2026. A promotional price has an end date by definition. Budgeting a year against a rate with a stated expiry is how a projected saving becomes a surprise increase in December.

The meter. This is the one almost nobody checks. Published OpenAI pricing carries a separate long-context meter where the current flagship tier moves from $5 to $10 per million input tokens and from $30 to $45 on output, with cached input doubling alongside the input rate. If your workload involves feeding in large documents, you may be paying the second set of numbers while planning against the first.

The channel. Reporting this week indicated Moonshot's Kimi K3 arriving on Amazon Bedrock at roughly 1.76 times OpenRouter pricing for the same model. One example rather than a rule, and enough to establish that where you buy is now a pricing variable alongside what you buy.

Cheaper is not the only direction

The same week also produced reporting that xAI's Grok 4.7 outperforms Claude on enterprise analysis while roughly doubling the bill. That is the pattern worth internalising: the spread between the cheapest and most expensive way to get a task done is widening, not collapsing toward a single falling price.

Which means the useful question stopped being which model is cheapest. It is which model is cheapest for the specific task, at the context length you actually send, through the channel you actually buy from. Three variables, and a headline rate answers none of them. We made the related argument about top-tier pricing in the best model now costs more.

What to do this week

Find your real cost per task. Not per million tokens. Take one workload you run regularly, look at an actual bill, and divide. That number is comparable across providers in a way that published rates are not.

Check whether you are on a promotional rate, and when it ends. Put the date in a calendar. This takes four minutes and prevents the most common AI budgeting surprise.

Check your context lengths against the meter threshold. If you routinely send long documents, you may already be on the higher tier. Trimming what you send is often cheaper than switching models, and it is the same discipline as cutting instructions that no longer earn their place.

Price your channel. If you buy through a cloud marketplace or reseller, compare against direct. Marketplaces can be worth a premium for billing consolidation, data agreements and procurement simplicity. That should be a decision, not an accident.

Keep switching cheap. With two flagship releases in a day, the model you standardise on today is unlikely to be the right one in six months. Abstracting your integrations so a swap is a configuration change rather than a project is the structural version of this advice, set out in hedging against model lock-in. The spend-control side is in turning on your spend controls.

Frequently Asked Questions