What custom AI chips mean for what you pay

Every few months a headline says a model lab is building its own chips. The current one is Anthropic, which confirmed in August 2026 that it has started an in-house silicon team [8]. If you run a small business on an API key and a couple of subscriptions, the question underneath the headline is short and practical: is my bill about to move, and in which direction.

The answer is that the chip story will not move your bill in any way you can act on, and the things that will move it are already sitting in your account settings. This guide is about telling those two categories apart, so you can stop reading infrastructure news as if it were a pricing announcement. It is not for anyone buying GPUs, negotiating a cloud contract, or training their own models. If your relationship with AI compute is a monthly invoice, this is the version that applies to you.

Custom silicon is a multi-year lever that never appears on your invoice

Start with what is actually confirmed, because the gap between that and the headline is most of the story. Anthropic’s careers page currently lists a Hardware team with openings for a Silicon Engineer and a Hardware Systems Architect, alongside a TPU Kernel Engineer role and a research engineering role in chip design [7]. TechRepublic reported on 6 August 2026 that Anthropic confirmed the team, that the role carries a salary of $320,000 to $485,000, and that applicants must show direct personal contribution to the completion and shipment of semiconductor designs [8]. Anthropic described the work as part of a broader multi-chip strategy that would complement rather than replace its existing use of hardware from AWS, Google, Nvidia and AMD, and did not disclose when a first custom chip could be ready [8].

That is it. A team is being hired. Anthropic itself has named no manufacturing partner, though the same report cites a June story from The Information saying the company was in talks with Samsung Electronics [8]. Designing a chip, taping it out, getting it manufactured and getting a datacentre’s worth of them running is a multi-year exercise even for companies that have done it before.

The deeper reason not to plan around it is structural. You do not buy compute, you buy tokens. No vendor itemises the hardware underneath, and none of them has ever published a per-chip cost. If custom silicon works exactly as intended, what you will see is a new model at a new posted price, and you will have no way to tell how much of that price came from the chip, how much from a change in model architecture, and how much from a decision about margin. Hardware news is a reasonable input to a five-year question about which lab survives. It is a useless input to a decision about what to run next Tuesday. Track posted prices, not fabs.

The tier you route work to moves your bill more than any chip will

Look at what a tier choice is worth today. Across the three main vendors, published input token prices run from $0.20 per million on OpenAI’s GPT-5.6 Luna to $10 per million on its GPT-6 Astra [5], a 50x range for the same nominal unit. Anthropic’s ladder sits inside that range: Claude Haiku 4.5 at $1 input and $5 output per million tokens, Claude Sonnet 5 at $2 and $10, Claude Opus 5 at $5 and $25, and Claude Fable 5.1 at $10 and $50 [1]. Google’s Gemini 3.8 Flash is $0.75 input and $3.75 output [6].

Two things follow. The first is that output tokens cost 5x input tokens at every Claude tier [1], so a prompt that asks for a summary and gets an essay is expensive in a way that is invisible if you only think about how much material you paste in. The second is that moving a task down one tier is a change you can make in an afternoon. Each step down the Claude ladder cuts that task’s token cost by 2x to 2.5x, on input and output alike: Opus 5 to Sonnet 5 is $5 down to $2 on input, and Sonnet 5 to Haiku 4.5 is $2 down to $1 [1]. At OpenAI the step from GPT-5.6 Terra at $2 input to GPT-5.6 Luna at $0.20 is 10x [5]. No plausible improvement in manufacturing gets close to that, and no manufacturing improvement is under your control.

The work of getting this right is not the routing, it is knowing which tasks tolerate it. Classification, extraction, tagging, first-pass drafting and reformatting usually survive a drop to the cheap tier. Anything where a subtle wrong answer is expensive and hard to spot usually does not. The only way to find out is to run 20 real inputs through both tiers and read the outputs side by side, which takes an hour and is worth more than any amount of reading about chips.

Prices drift down at a tier, but a posted price is not a promise

There is a real trend and it is in your favour. Claude Sonnet 4.5 and Sonnet 4.6 are priced at $3 input and $15 output per million tokens; their successor, Claude Sonnet 5, is $2 and $10 [1]. That is roughly a third off for the same position in the range, and it is the pattern that has held across generations. Waiting is often the cheapest optimisation available.

The trend is not a guarantee, and two live examples on vendors’ own pricing pages make the point. Gemini 3.8 Flash holds its $0.75 input and $3.75 output pricing through 31 December 2026, then doubles to $1.50 and $7.50 on 1 January 2027 [6]. OpenAI lists GPT-5.6 Sol at $4 input and $20 output, with that promotional pricing available at least through 21 November 2026 [5]. Neither is hidden. Both are dates that will arrive.

If your margin on a service depends on a promotional rate, it depends on a date. Write that date in the same place you keep renewal dates for anything else, and rerun the arithmetic before it passes rather than after the invoice arrives. Vendors also add surcharges for particular modes of use, which are easy to switch on and forget: OpenAI charges double for its fast processing mode, and a 10% uplift on regional data residency endpoints for models released on or after 5 March 2026 [5].

The two discounts you already control beat anything a fab will hand you

Prompt caching is the first. Anthropic charges cache reads at 0.1x the base input price on most models, and 0.025x on Claude Fable 5.1 and Mythos 5.1, with writes at 1.25x for a 5-minute lifetime or 2x for a 1-hour lifetime [1][2]. On Sonnet 5, that turns the repeated part of your prompt from $2 per million tokens into $0.20 [1][2]. The constraints are specific: a cache hit requires 100% identical prompt segments up to and including the block you marked, the cached content has to sit at the start of the prompt, the minimum cacheable prefix ranges from 512 to 4,096 tokens depending on the model, and you get at most 4 cache breakpoints per request [2]. A cache write pays for itself after a single read on the 5-minute lifetime and two reads on the 1-hour one [2].

Batching is the second. Anthropic’s Message Batches API is 50% off both input and output, with most batches finishing in under an hour, a hard expiry at 24 hours, and a limit of 100,000 requests or 256 MB per batch [3]. Results stay retrievable for 29 days [3]. The two discounts stack, though cache hits inside a batch are best-effort rather than guaranteed [3].

None of this is Anthropic-specific, which is the useful part. OpenAI prices cached input tokens 90% below standard input and gives 50% off for batch [5]. Google discounts context caching by 90% and batch by 50% [6]. Whatever you build, the same two levers exist at whichever vendor you move to. The cost of using them is architectural rather than financial: your system prompt has to be stable and go first, and some of your work has to tolerate an hour of latency instead of a second.

Model retirement is the cost event that actually reaches you

The pricing risk most small operators worry about is a price rise. The one that actually costs them a weekend is a retirement. Anthropic commits to at least 60 days’ notice before retiring a publicly released model, and publishes the schedule [4]. Recent entries show claude-opus-4-1-20250805 retired on 5 August 2026, and claude-sonnet-4-20250514 and claude-opus-4-20250514 both retired on 15 June 2026 [4]. After a retirement date, requests to that model fail [4].

Migrations are not always drop-in either. The temperature, top_p and top_k parameters are deprecated for Claude Opus 4.7 and later, and return a 400 error if set to anything other than the default [4]. A workflow tuned around a specific temperature setting does not just get a new model name, it needs the tuning moved into the prompt.

Two habits make this a Tuesday afternoon rather than an emergency. Keep the model name in one configuration value rather than scattered through your prompts, and keep 20 to 30 real inputs with outputs you know are good. When the deprecation email arrives, you swap the string, run the set, and read the diffs. Without the set, you are testing by vibes on someone else’s deadline.

Measure cost per finished job, not cost per token

Token prices are the wrong denominator for almost every decision a small operator makes. What you care about is the cost of one finished thing: one processed invoice, one drafted reply, one summarised call. That number includes retries, includes the times the cheap model got it wrong and you ran it again on the expensive one, and includes the minutes you spent fixing output.

Once you track it, the comparisons get honest. A tier that is cheaper per token but needs two attempts and a manual edit is not cheaper. A caching setup that cuts the repeated prefix to a tenth of its price [2] is worth exactly nothing if your prompt template changes on every request and never hits the cache. And a vendor’s new generation at a lower posted price [1] is only a saving if your prompts still work on it.

This is also the number that makes the chip story readable. If the industry’s hardware efforts eventually cut inference costs, it will reach you as posted prices at some tier falling again, the same way they already have. You will notice it in your cost per finished job a quarter later, and there will be nothing to do about it except keep taking the cheaper tier when it passes your evals.

What still goes wrong

The biggest limit is that you cannot verify any of the upstream story. Vendors do not publish cost per token to serve, so you cannot tell whether a price cut reflects cheaper hardware, a smaller model, a change in margin, or a decision to buy market share. You can only read posted prices, and posted prices are a commercial decision that happens to be downstream of hardware. Anyone who tells you a specific chip programme will produce a specific discount for you is guessing.

The discounts have sharper edges than their headline percentages suggest. Cache hits inside batches are best-effort, with reported hit rates ranging from 30% to 98% depending on traffic patterns [3], so a bill that assumes both discounts apply in full will come in higher than modelled. Prompts shorter than the model’s minimum are simply not cached, and no error is raised [2], which means you can believe caching is on for weeks while paying full price. And the 24-hour batch expiry [3] is a real operational risk for anything with a deadline attached.

Finally, the cheap-tier discipline this guide recommends has a failure mode that the invoice hides completely. A weaker model that produces plausible, wrong output costs you nothing extra in tokens and quite a lot in the work of catching it, or the work of not catching it. Cost per finished job only captures that if you are honest about the time spent checking. If you are not going to check, stay on the tier you trust and spend the difference.