Claude Haiku 5.5 lands 90% cheaper, Broadcom lines up more than $50bn for OpenAI's chips, and 3 more AI moves
Five verified developments from October 7 to 8, 2026: Anthropic prices Claude Haiku 5.5 at 90% below its predecessor, OpenAI puts GPT-6 in front of every ChatGPT tier with an interface the model writes, Broadcom and Oracle line up more than $50bn of private debt for OpenAI's chips, Isomorphic Labs is reported in talks at a $40bn valuation, and the Fed's September minutes name the AI buildout as an inflation driver.
In brief
Between October 7 at 12:00 UTC and October 8 at 12:00 UTC, Anthropic cut the price of its small Claude model by 90%, OpenAI began putting GPT-6 in front of every ChatGPT user with an interface the model writes for itself, Broadcom and Oracle moved to borrow more than $50 billion against chips, Isomorphic Labs was reported to be raising at a valuation of at least $40 billion, and the Federal Reserve's September minutes named the AI buildout as a reason goods prices keep climbing.
- Launches: Anthropic released Claude Haiku 5.5 at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, a rate the company says is 90% below Haiku 4.5 for those requests. It is the same rate OpenAI lists for GPT-6 Luna, so the two now sit together at the bottom of the price table among models from frontier labs. Anthropic also halved Sonnet 5.5's cache-read price and began paying monthly API credits to Max and Team subscribers. 12
- Launches: OpenAI began a global rollout of GPT-6 in ChatGPT on October 7 for the Plus, Pro, Business and Enterprise tiers, reaching Free and Go on October 8, with a capability it calls Intelligent UI that builds an answer out of charts, buttons, forms and small interactive tools. Paid tiers run on GPT-6 Sol and the free tiers on GPT-6 Luna, and OpenAI puts the audience at more than 1.2 billion weekly users. The models behind ChatGPT Work and Codex are unchanged by the rollout. 34
- Funding: Broadcom is arranging more than $50 billion of financing for the custom chip it is developing with OpenAI, with Apollo Global Management and Blackstone among the lenders in talks, and Oracle is negotiating with Apollo and Goldman Sachs to fund a chip purchase of its own, according to The Wall Street Journal. The Journal's account, relayed by Bloomberg and by Investing.com, describes structures in which private investors fund an entity that buys the chips and leases them back, so the buyer takes on the lease while the hardware sits with the investor. Broadcom has borrowed for chips before: the platform it set up in June with Apollo and Blackstone as anchor investors planned more than 20 gigawatts of capacity through 2028, starting with a $35 billion tranche. 56
- Funding: Isomorphic Labs, the drug-discovery company spun out of Google DeepMind, is in early talks to raise money at a valuation of at least $40 billion and possibly as high as $50 billion, according to people familiar with the effort. The round would come five months after a $2.1 billion Series B led by Thrive Capital. Isomorphic has partnerships with Eli Lilly and Novartis and has said it is working on cancer and immune disorders. 78
- Policy: The minutes of the Federal Reserve's September 15-16 meeting, released on October 7, record officials attributing upward pressure on prices to "surging AI-related investments" alongside higher crude oil and fuel prices caused by the Iran conflict. The staff projected inflation stepping down to 2% in 2029 as the effects of tariffs, geopolitics and the AI buildout fade, and the minutes name heavy private debt issuance for AI infrastructure as a factor in higher long-term Treasury yields. 910
Five developments
1. Anthropic puts the small-model floor at ten cents
Anthropic released Claude Haiku 5.5 on Wednesday, calling it the cheapest, fastest and most capable small model it has shipped, and priced it at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. Above that threshold the same model costs $0.50 and $2.50. Haiku 4.5, the model it replaces in that slot, listed at $1.00 and $5.00, so the new rates are 90% lower for short requests and 50% lower for long ones. Anthropic says about 90% of Haiku 4.5 requests fell under the 100,000-token line, and estimates that a typical workload now costs around 75% less to run, a figure that accounts for the new tokenizer, which spends more tokens on the same work. 12
The rate is the news, because it lands exactly on OpenAI's. GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output tokens as base rates, with a surcharge beginning above 272,000 input tokens, which VentureBeat confirmed against OpenAI's published pricing on October 7. Google's Gemini 3.5 Flash-Lite lists at $0.30 and $2.50, Gemini 3.8 Flash at $0.75 and $3.75 on promotional rates that run through December 31, and Grok 4.3 at $1.25 and $2.50. The step up to the next tier is large: Anthropic's own Sonnet 5.5 and OpenAI's GPT-6.1 Sol both list at $2.00 and $10.00 per million tokens. 2
What a job costs depends on how many tokens it takes, and Haiku 5.5 shows why. The independent evaluator Artificial Analysis scored the model at 43 on its Intelligence Index when run at maximum effort, against a median of 13 for models in the same price band, and measured 243 output tokens per second. It also recorded 440 million output tokens across the index, against a median of 100 million, which the evaluator describes as very verbose: a chatty model at a low per-token price can still cost more per task than a terse one. Artificial Analysis puts the cost at $0.21 per index task. 11
Anthropic's own benchmark table claims a large jump over the model it replaces, with Sonnet 5.5 still ahead on the hardest work. On GDPval-AA v2.1, which scores agentic professional work, Haiku 5.5 rates 1,620 against 735 for Haiku 4.5, 1,437 for GPT-6 Luna and 1,840 for Sonnet 5.5. On AA-Briefcase v1.1 it scores 1,578 against 614, 1,336 and 1,824 respectively. On OSWorld 2.1's offline subset, a test of operating a real computer, it reaches 72.4% against 15.7%, 48.9% and 83.9%. On Terminal-Bench 4.0 it scores 39.2% against 0.0%, 16.4% and 70.6%. 12
The effort setting changes what those numbers mean. Haiku 5.5 is the first Haiku-class model with an adjustable effort control, and its default is medium; the 39.2% on Terminal-Bench 4.0 is the maximum-effort result, and the medium setting scores around 20% on the same benchmark. Anthropic's framing is that the larger models remain the right choice for complex coding agents and that Haiku belongs on narrowly scoped work — compaction, summarisation, lookups, one step of a larger job. The company cites Rogo, where a Haiku subagent pulls a segment revenue line out of a filing while a bigger model builds the deck. 12
Two changes ship alongside the model. Cache reads on Sonnet 5.5 fall from $0.20 to $0.10 per million tokens, which Anthropic says cuts the cost of most agentic work on that model by about 20%. Max subscribers get $100 a month in API credits on the 5x plan and $200 on the 20x plan, and Team subscribers get up to $500 pooled across their users. Anthropic's Python and TypeScript SDKs add support for browser use and computer use in beta, tasks it says Haiku 5.5 suits because of its combination of speed and price. 1
Haiku 5.5 is available now on Anthropic's own platform, Amazon Web Services, Google Cloud and Microsoft Azure under the identifier
claude-haiku-5-5, reads text and images and writes text, and carries a 1 million token context window. Its cybersecurity safeguards are more restrictive than Haiku 4.5's, and they block penetration testing; organisations that need wider cyber or biology access can apply to Anthropic's verification programmes. 1112. GPT-6 becomes the default, and the interface becomes model output
OpenAI began rolling GPT-6 out inside ChatGPT on October 7, first to the Plus, Pro, Business and Enterprise tiers and from October 8 to the Free and Go tiers. Paid tiers are served by GPT-6 Sol and the free tiers by GPT-6 Luna, the model whose $0.10 and $0.50 price Anthropic matched the same day, and OpenAI describes the audience as more than 1.2 billion weekly users. The models behind ChatGPT Work and Codex are unchanged, and enterprise availability depends on each workplace's admin settings. 34
The new capability is called Intelligent UI, and it lets a reply contain graphics, tappable buttons, forms, charts and interactive experiences rather than text alone. OpenAI's own examples: asking how a seven-speed bicycle is put together returns a diagram whose five systems can be selected one at a time, and asking for a Sunday roast plan returns images of the dishes, a shopping-quantity calculator that adjusts to the number of guests, and a checklist timeline. Users can also ask for a tool for the moment — a savings calculator, a bill splitter, a game playable inside the conversation. 34
OpenAI built a library of native components and a compiler that processes the interface as the model generates it, so the elements appear progressively instead of arriving with the finished answer. The company says it expanded its training methods so the model decides on content, layout, visuals and interaction, and that it evaluated the interfaces it produced for clarity, usefulness and completeness. It also says the model's design judgment still has room to improve. 3
Answer speed changes with the same release, and these figures come from OpenAI's internal testing; no independent run of the same comparisons has been published. The model can now begin answering while it continues to think, interleaving partial responses; on questions that need a web search, OpenAI says GPT-6 Instant starts answering 44% sooner on average than GPT-5.6 Instant. In its internal evaluation of everyday agentic tasks, GPT-6 Extra High begins answering in the same time as GPT-5.6 Medium and scores better overall than GPT-5.6 Extra High. 34
OpenAI says GPT-6 builds on safety work done for Astra, resists multi-turn attempts to bypass its training more strongly than GPT-5.6 Sol, and is better at stating the limits of what it can do; a system card accompanies the release. What a reader can act on today is narrower than the announcement: the September launch put GPT-6 Sol and Luna on the API at $2.00/$10.00 and $0.10/$0.50, and this release moves the same family into the default chat surface for paid and free users alike, with an interface the model composes at generation time. 3
3. Broadcom and Oracle reach for private credit to buy chips
Broadcom is arranging more than $50 billion in financing for the custom AI chip it is developing with OpenAI, with Apollo Global Management and Blackstone among the lenders in talks to take part, and Oracle is negotiating with Apollo and Goldman Sachs to line up money for a chip purchase of its own. The Wall Street Journal reported both on October 7, citing unidentified people familiar with the discussions; Bloomberg relayed the report at 21:25 UTC, and The Business Times carried the same Bloomberg text later that evening. 51213
The structure matters more than the headline number, and the relays describe Oracle's version in the most detail. Private investors would fund an off-balance-sheet entity that buys the chips and leases them back to Oracle, which expands a cloud provider's capacity while keeping the hardware off its own balance sheet. Investing.com, summarising the Journal's report, says the talks are at an early stage and that the parties are aiming at deadlines around the end of the year. 6
The precedent is Broadcom's own. In June the company said it had set up a platform, with Apollo and Blackstone as anchor investors, to help finance more than 20 gigawatts of computing capacity for AI developers including Anthropic and OpenAI through 2028; it opened with a $35 billion tranche led by Apollo in partnership with Blackstone. Bloomberg reported in August that Broadcom had held talks with lenders to raise more than $60 billion for a chip-financing deal benefiting Anthropic and other companies. 513
The list of lenders is the same short list in every deal. Apollo, Blackstone and Goldman Sachs are the names appearing across the Broadcom package, Oracle's chip financing and, as reported on October 6, the $40 billion SpaceX is seeking to buy Nvidia chips. The private credit funds are becoming the standing counterparty for AI hardware, which lets a chip buyer pay over years instead of up front and keeps the resulting debt off its reported balance sheet. 56
Everything in this section rests on people the Journal does not name. None of Broadcom, Oracle, OpenAI, Apollo, Blackstone or Goldman Sachs has confirmed the amounts or the structures in the accounts available here, and the talks could end without a deal. Treat the figures as reported intentions under discussion, with a close still ahead. 56
4. Isomorphic Labs is reported in talks at $40 billion and up
Isomorphic Labs, the drug-discovery company spun out of Alphabet's Google DeepMind, is in early talks to raise new funds at a valuation of at least $40 billion, and one person familiar with the effort put the figure as high as $50 billion. Bloomberg reported the talks on October 8 at 4:00 UTC. The financing is still open, the terms could change, and Isomorphic had no comment when the news organisation asked. 7
The company is London-based and was formed in 2021 to build on AlphaFold, the protein-structure model that made DeepMind's reputation in biology. Demis Hassabis, who moved from running DeepMind to a chairman role there and to Alphabet's chief scientist post in August 2026, remains Isomorphic's chief executive. The company announced partnerships with Eli Lilly and Novartis, has said it is focusing on cancer and immune disorders, and is aiming to design and develop drugs in house as well as work with pharmaceutical partners on its AI drug design engine. 7
The baseline for judging the number is five months old. In May, Isomorphic raised $2.1 billion in a Series B led by Thrive Capital, with Alphabet and GV returning alongside new investors MGX, Temasek, CapitalG and the UK Sovereign AI Fund. The company left the post-money valuation of that round undisclosed, which leaves the jump to a $40 billion talk without a published starting point to measure it against. 78
The sector it sits in supplies the caution. Bloomberg notes that while AI developers promise to cut the time and cost of drug development, few AI drug-discovery companies have progressed to clinical trials and none has brought a medicine to market. Hassabis has said his goal is to shorten development from years to months. A $40 billion valuation would put a number on that promise, ahead of any approved product. 7
Isomorphic is also competing with its own industry's model makers. OpenAI released a model aimed at life-sciences research this year, which puts a general-purpose lab into the same territory as a company built specifically to apply AI to drug design. 7
5. The Fed writes the AI buildout into its minutes
The Federal Reserve released the minutes of its September 15-16 policy meeting on October 7 at 2:00 p.m. Eastern, and the AI buildout appears in them as an economic force, with officials treating it as part of the inflation picture. Officials cited "surging AI-related investments" alongside higher prices for crude oil and refined fuel caused by geopolitical developments as contributors to inflation pressures. 910
The minutes record the reasoning in the participants' own words. Several participants observed that core goods prices also kept rising, "as effects of the AI buildout appeared to increase while the effects of tariff increases waned." Some commented that the buildout could push aggregate demand past what the economy can supply over the medium term, putting upward pressure on inflation, and some pointed to higher transport and input-material costs at businesses affected by it. A couple of participants argued that a higher policy rate would help stop price increases tied to AI demand from broadening into more persistent inflation. 14
The staff forecast published with the minutes has inflation stepping down over the next two years and reaching the Fed's 2% target in 2029, as the effects of tariffs, geopolitical developments and the AI buildout are expected to wane. The same staff projection attributed a share of recent core inflation to "an increase in technology-related consumer goods prices associated with the AI buildout," and the staff described the uncertainty around the outlook as substantial, naming the potential economic effects of AI investment and adoption alongside stubbornly elevated inflation and geopolitics. 14
One line connects the document to the borrowing reported the same week. The minutes list "competition for capital from heavy private debt issuance to finance the development of artificial intelligence (AI) infrastructure as also contributing to higher term premiums and Treasury yields," which is the Federal Reserve describing the debt financing of data centres as something that moves the price of government borrowing. Several participants said the scale and pace of the buildout had continued to surprise to the upside, and some observed that strong demand for skilled workers in AI-related sectors was driving strong wage gains for those workers. 14
Participants generally expected AI investment to produce stronger productivity and output growth in the coming years, while saying the size and timing of those gains remain uncertain. The Washington Post, which reviewed recent minutes, reported that AI barely came up in Fed discussions a year ago and has become a recurring subject this year. 1014
The read-through
Three of this issue's five items are about who ends up owning the hardware, and the money behind it. Broadcom is assembling more than $50 billion of private debt to finance the custom chips OpenAI is buying, and Oracle is negotiating a structure in which outside investors own the chips and lease them back. Isomorphic Labs is reported to be raising equity at a valuation of at least $40 billion, five months after a $2.1 billion round whose valuation went undisclosed. Both are the same appetite in different instruments: capital committed today for capacity that is still to be built. This reading is this issue's own inference from separate reports; each company described its own financing on its own terms. 567
The Fed's minutes supply the link the individual financing stories leave out. The document records officials attributing price pressure to "surging AI-related investments", and the same set of minutes names heavy private debt issuance for AI infrastructure as a contributor to higher term premiums and Treasury yields. Those two lines describe the same spending Broadcom and Oracle are borrowing against, which makes the terms they negotiate a matter beyond the two companies: a wider spread on this debt raises the cost of the money behind the whole buildout. 14
The counter-current is running at full speed on the product side. Anthropic cut Haiku 5.5 to $0.10 and $0.50 per million tokens for short requests, exactly matching OpenAI's GPT-6 Luna, and OpenAI's own release puts that same Luna in front of every free ChatGPT user. The price of a unit of model work is falling while the price of the capital that funds the chips may be rising, and those two movements have different clocks: inference prices can be cut in an afternoon, while debt has to be refinanced. A team budgeting for either one this week should size its assumptions against the other. 13
The last item is a caution about a number. Isomorphic's reported $40 billion to $50 billion valuation would be set on a company that has partnerships with Eli Lilly and Novartis and no medicine on the market, five months after a round that raised $2.1 billion without publishing a price. The valuation is a price on the promise of AI-designed medicine, set ahead of any AI-designed medicine reaching a patient. 78
On sourcing, the issue divides cleanly. The Fed's minutes are a primary document released by the agency, and the quotations above are taken from it; the Washington Post added the context that AI has become a recurring subject in recent minutes. The Broadcom and Oracle figures come from a Wall Street Journal report that relies on unidentified people, relayed here through Bloomberg and Investing.com, with no company confirming the amounts or the structures. The Isomorphic valuation comes from people familiar with the talks, in a round that is still open. Anthropic's benchmark table is the vendor's own measurement, and the exception in this issue is Artificial Analysis's index score, its throughput figure and its cost per task, which the evaluator measured itself; the GPT-6 speed and quality comparisons are OpenAI's internal evaluations. 5101114
Watch next
- Whether the ten-cent floor holds. Haiku 5.5 and GPT-6 Luna now list at the same input and output prices, and Google's Flash tiers sit above them. A response from Google, or a further cut from either lab, would show how much of the small-model tier's margin is left. Watch the expiry of Gemini 3.8 Flash's promotional rates on December 31 as the first scheduled test. 2
- Whether the GPT-6 rollout completes on schedule and what Intelligent UI does to usage. The Free and Go tiers were promised GPT-6 Luna from October 8, with enterprise access gated by admin settings. Independent measurement of GPT-6 Luna's cost per task, and any shift in how long people stay in the app, would test whether a generated interface changes usage or only the look of an answer. 3
- Whether the Broadcom and Oracle structures close, and on what terms. Both sets of talks are reported as early stage, with deadlines around the end of the year. A completed financing, a rating assigned to the debt, or a deal that quietly ends would each say something different about how much of the buildout private credit is prepared to carry. 6
- Where Isomorphic's round prices. The reported range runs from $40 billion to $50 billion, and the May Series B left no published valuation to anchor it. A closed round at either end would set the first public reference price for an AI drug-discovery company at this scale. 7
- Whether the AI language stays in the next set of minutes and whether it reaches the data. Officials have now written AI investment into their account of goods prices and long-term yields while projecting inflation back to 2% in 2029 as its effects fade. The next minutes, and the inflation and productivity prints between now and then, are where that expectation either starts showing up or gets revised. 1014
References
- 1Anthropic — Claude Haiku 5.5
anthropic.com
- 2
- 3
- 4
- 5
- 6
- 7
- 8Isomorphic Labs — Isomorphic Labs announces Series B investment round
isomorphiclabs.com
- 9
- 10
- 11
- 12
- 13
- 14
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.