Funding is not a shipping date

You read a headline about an AI company raising money, and some part of your brain files it as progress. Nine figures, a name you half recognise, a valuation with a “b” in it. The implicit promise is that the thing described will exist soon, and that you will be able to use it. Most of the time the funding number tells you almost nothing about either. It tells you the company can pay salaries for a while. Whether it can build the thing, at volume, on the date implied, is a separate question with a separate answer, and that answer is usually published somewhere much duller than the press release.

This matters to a one-to-ten person business in a specific, unglamorous way. You are not deciding whether to buy a robot. You are deciding whether to wire six weeks of your own work into a tool, and you need to know how long that tool will be there in the form you built against. This guide is about reading vendor announcements for the constraint rather than the number, and about the two constraints that actually bite small operators: capacity you are allowed to use, and the shelf life of the specific model version your automations name. It is not a procurement guide for industrial hardware, and it will not tell you which vendor to pick.

A funding round buys time, not units

Take a case with an unusually clear paper trail. Humanoid, a London robotics company [8], announced on 21 May 2026 that Robert Bosch would act as its contract manufacturing partner for HMND 01 robots destined for the European market, following a proof of concept in March 2026 at Bosch’s plant in Bühl, where the robots autonomously moved boxes from a conveyor onto trolleys across five box sizes of differing footprint, height and weight [7]. Two months later, on 21 July 2026, the company announced a $152 million Series A at a $1.35 billion post-money valuation, bringing total funding to $270 million [6]. Bosch’s chief technology officer described its role as contract manufacturing plus strategic consulting and technical expertise in hardware design, production and supply chain [6].

The headline number is the $152 million. The number that tells you when robots exist is somewhere else. Reporting on the same announcements puts a beta phase in late 2026, a gamma phase in 2027 and a mass-production version beginning in 2028, with the Schaeffler agreement covering a four-digit number of wheeled units across Schaeffler’s global facilities by 2032 and first rollouts running from December 2026 to June 2027 at two German sites [8]. Read those two paragraphs together and the shape of the industry is plain. Capital arrived years before units will. The company’s own release calls the Schaeffler agreement the industry’s largest publicly announced commercial agreement of its kind [6], and it is measured in thousands of robots over six years, not hundreds of thousands next quarter.

Nothing here is a criticism of the company. Building precision hardware at volume is slow, and borrowing an established manufacturer’s production lines instead of building your own is the sensible move. The point is narrower and it generalises: the announcement was about money, the constraint was manufacturing, and the two were reported in the same breath as though they were the same fact. When you see a large round with no named manufacturer and no dated production phase, you are looking at a company that still has to solve the harder half after the money lands.

The software version of a factory is a retirement date

You will probably never buy a humanoid robot. You will absolutely build something on a model name, and the software equivalent of “who builds this and when” is “how long does this stay available, and how much warning do I get.” That is not a matter of speculation. All three major vendors publish it.

Anthropic states that it gives at least 60 days’ notice before retiring a publicly released model, and publishes a table of every model with its state and a tentative retirement date [1]. As of today that table shows claude-sonnet-4-5-20250929 retiring not sooner than 29 September 2026, claude-haiku-4-5-20251001 not sooner than 15 October 2026 and claude-opus-4-5-20251101 not sooner than 24 November 2026, alongside models already gone: claude-3-7-sonnet-20250219 retired on 19 February 2026, and claude-opus-4-1-20250805 deprecated on 5 June 2026 and retired on 5 August 2026 [1]. OpenAI’s policy is longer at the top end and shorter at the bottom: at least 6 months for generally available models, at least 3 months for specialised variants of them, and preview models may be retired with much shorter notice, such as 2 weeks [4]. Its published schedule shuts down gpt-3.5-turbo-0125, gpt-4-turbo and o1-2024-12-17 on 23 October 2026, and gpt-5-2025-08-07 on 11 December 2026, naming gpt-5.6-terra and gpt-5.6-sol as the recommended replacements [4]. Google’s Gemini documentation is blunter still: preview models are deprecated with at least 2 weeks’ notice, and for the “latest” aliases, a 2-week email notice is what you get before the version behind the alias changes underneath you [5].

So the same three-tier reality applies to software as to hardware. There is the thing announced, there is the thing generally available, and there is the preview thing that may not survive the month. A preview model with a 2-week notice period is the software equivalent of a robot with a 2028 mass-production date. It is real, it works, and it is not something to build your invoicing on.

Pin the version, then put the retirement date in your calendar

The practical consequence is small and boring, which is why almost nobody does it. If your automation calls a model by a floating alias, you have handed the vendor permission to change your output without changing your code, and Google says explicitly that a 2-week email is what you get in exchange [5]. If it calls a dated version like claude-sonnet-4-5-20250929 or gpt-5-2025-08-07, the behaviour holds still, and the vendor’s published table tells you the date it stops [1][4].

Pinning is the easy half. The half people skip is the calendar entry. Take the retirement date from the vendor’s table, subtract a month, and put a reminder there with a link to the automation it affects. Sixty days of notice is genuinely enough time if the notice reaches a human who knows which Zapier step or n8n workflow is affected. It is not enough if it lands in a billing inbox nobody opens, which is easy to arrange without meaning to when the account was set up eighteen months ago and the work has moved on since.

When the notice does arrive, the work is re-testing rather than rewriting. Move the pinned version forward, run the same ten real inputs you used when you built the thing, and compare. If you did not save those ten inputs, that is the actual cost of the migration, because you will be reconstructing your own acceptance criteria from memory while under a deadline set by someone else.

What you are allowed to use, and what it costs

Capacity is the other constraint that gets left out of announcements, and it is published too. Anthropic’s rate limit documentation lists usage tiers with monthly spend caps: $500 at Start, $1,000 at Build, $200,000 at Scale, and a custom tier above that with no monthly cap [3]. The published per-model limits for Claude Opus 5 and Claude Sonnet 5 are 1,000 requests per minute, 2,000,000 input tokens per minute and 400,000 output tokens per minute [3]. For a business your size the throughput ceiling is irrelevant and the spend cap is not. If you are on a tier with a $500 monthly cap and you build something that a client’s traffic pushes past it, the ceiling you hit first is the billing one.

One capacity detail is worth knowing because it changes design decisions. Tokens read from cache do not count toward the input-tokens-per-minute limit on most models [3], and cache reads bill at 0.1x the base input rate [2]. That means a long, stable system prompt you reuse all day is cheap and does not eat your throughput, while the same text pasted fresh into every call is neither.

Price is the number most likely to move, and it does not only move down. Claude Opus 5 lists at $5 per million input tokens and $25 per million output, Claude Sonnet 5 at $2 and $10, and Claude Haiku 4.5 at $1 and $5 [2]. Sonnet 5 undercuts the Sonnet 4.5 it follows, which lists at $3 and $15 [2], but newer does not mean cheaper across the range: Claude Fable 5.1 lists at $10 and $50 [2]. Batch processing takes 50% off both sides [2]. Google, meanwhile, lists Gemini 3.8 Flash as its newest stable model while Gemini 3.1 Pro is still marked preview [5], which is the three-tier reality again in another vendor’s words. The judgement worth making is not which model is cheapest this month. It is that price is a thing you can renegotiate or route around, while availability and notice periods are the things that fall over without warning. Optimise your attention accordingly.

Four questions that work on any vendor, hardware or software

The Humanoid case and the deprecation tables are the same question asked in two industries, so the same four questions work on both. Who actually builds or serves this, by name. At what volume or capacity, as a published number rather than an ambition. What is the notice period before it changes or goes away, in writing. And what specifically happens to my work when it does.

A vendor that answers all four with links is telling you something different from one that answers with a valuation. In software, all four answers already exist as public documentation pages, which means a vendor that cannot point you at them is a vendor whose product is thinner than its marketing. In hardware, the equivalent of those documentation pages is a named manufacturer and a dated production phase, which is exactly what the Bosch arrangement supplies and what the funding round does not [7][8].

The asymmetry worth internalising is that good news is announced and bad news is published. Rounds, partnerships and benchmark wins arrive as press releases in your feed. Retirements, rate limits and spend caps sit on documentation pages you have to go and read. If your picture of a vendor comes entirely from the first category, it is systematically optimistic, and the correction is fifteen minutes on the pricing, limits and deprecation pages before you commit a week of work.

What still goes wrong

Published policies are floors, not promises, and floors have trapdoors. Anthropic’s 60-day commitment covers publicly released models, and its own definition of a deprecated model is one that is still functional but no longer recommended, with the warning that deprecated models are likely to be less reliable than active ones [1]. A model can therefore degrade in practice well before it disappears on paper. OpenAI’s notice periods carry the same kind of escape hatch, applying unless safety or compliance concerns require a faster timeline [4]. Neither vendor owes you the old behaviour of a model that is technically still running, and “the outputs got worse” is not a retirement event you will be notified about.

The four questions also do not protect you from the failure mode where everything works and the economics do not. A vendor can ship on time, honour every notice period, and still price the tier you depend on out of your reach, or fold the feature into a plan that costs ten times what you pay now. Published prices are current prices, and nothing on a documentation page is a commitment about next year.

And the robotics half of this comes with an honest caveat: none of it is likely to touch a small business directly this decade. A four-digit unit count spread across a multinational’s global facilities by 2032 [8] is not a supply curve that reaches a five-person company. The reason to follow it at all is that it is the same reasoning problem you face every time you choose a tool, run at a slower speed and with a clearer paper trail. If you want to practise reading the constraint instead of the headline, hardware is where the gap between the two is easiest to see.