The Read

The price of using intelligence and the price of financing it moved in opposite directions inside the same forty-eight hours. DeepSeek pushed a quiet update to its Flash model that now posts 77.0 on DeepSWE v1.1 — ahead of Claude Opus 5 at 74.0 and GPT-5.6 Sol at 73.0 — serves at $0.15 and $0.60 per million tokens off-peak — $0.30 and $1.20 at peak — and opened its own web search and fetch endpoints at $0.75 per thousand requests. Separately the Financial Times reported — single-outlet, and Reuters could not immediately verify it — that Anthropic has offered Blackstone a coupon above 10% on $9B of an $18B junior tranche inside a $60B Broadcom-arranged package, against roughly 7% for comparably rated public bonds. Both facts are the same fact: inference is becoming a commodity priced in fractions of a cent while the capacity to produce it gets financed like a toll road. The day-zero read is unchanged and sharper for it — the opening is not buying tokens more cheaply, it is finding the old problems that were never worth solving at the old price, because the price keeps moving to places an incumbent's cost base cannot follow.

🌊 Tide

No shift. One confirmation. cost-collapse gets its cleanest challenger-side datapoint since the summer, and the mechanism is the KV-cache compression work DeepSeek published in September rather than a price war: a model whose vendor serves it from $0.15 per million input tokens leads a $5-input frontier model on the agentic-coding benchmark its vendor leads with, and the same vendor priced live web retrieval like bandwidth. The counter-beat belongs to the waves, not the tide — the cost of the capital behind the capacity is rising even as the cost of the output falls, and a 10% coupon on compute debt is not evidence against the tide. It is evidence that somebody has to pay for the fab.

An off-peak $0.15 model passed Opus 5 on DeepSWE, and its vendor priced web search like bandwidth

DeepSeek posted DeepSeek-V4.1-Flash-1017 to Hugging Face on Saturday morning UTC, MIT-licensed, described on its own model card as "a minor update" to V4.1-Flash. The numbers are not minor. On DeepSWE v1.1 it scores 77.0 against 74.2 for the model it replaces, 74.0 for Claude Opus 5 and 73.0 for GPT-5.6 Sol. Terminal-Bench 4.0 goes 31.2 to 42.3 — past GPT-5.6 Sol's 39.9 and GLM-5.3's 37.9, still well behind Opus 5's 51.8. Apex Bench v1.1 goes 64.6 to 72.1, Automation-Bench to 60.5, Agents' Last Exam to 35.4. On raw reasoning it stays behind: HLE without tools 42.6 against Opus 5's 56.3. All eleven published benchmarks improve on the prior Flash, and all eleven are DeepSeek's own figures with its own harness choices, which is where the usual discount applies. Note also which models the card leaves out of its comparison table: on self-reported DeepSWE v1.1 numbers, Gemini 4 Argon sits at 77.9 and Meta's Muse Wonder pair above that, so 77.0 is not the field lead — it is a pass on the frontier model most teams are actually paying for. The architecture underneath is not new on Saturday and is worth understanding anyway: the V4.1-Flash technical report from September, titled "Pushing the Limits of KV Cache Compression," puts global KV at 890 bytes per token under FP4 main-KV quantisation with cross-layer reuse, roughly a quarter of V4-Flash, with a causal encoder-decoder split that nearly halves prefill compute and 8B active parameters in prefill against 16B in decode. The model card calls the update minor; the changelog confirms the calling method and pricing are unchanged — $0.15 input, $0.60 output and $0.003 per million cache-read tokens off-peak, doubling to $0.30, $1.20 and $0.006 during peak hours. Alongside it, DeepSeek opened Search and Fetch endpoints on its own API — ranked web results with site:, lang:, after: and before: operators, pages returned as Markdown — at $0.75 per thousand successful requests, 5 QPS per account, failures and empty results free, and no charge for server-side calls routed via DeepSeek Harness, Claude Code or Codex until billing starts at 16:00 UTC on 31 October.

So what: Two things open here. The first is a cheap, downloadable, genuinely competent coding agent you can run or rent without a procurement cycle — benchmark it on your own harness this week, because the gap between a vendor's agentic numbers and your repository is where the real decision lives. The second is quieter and bigger: when a frontier lab starts selling retrieval at $0.75 per thousand calls, the retrieval layer of your stack stops being a build decision. Price your own search plumbing against that line — and against the peak rate, not the headline one, and if it does not clear, spend the engineers on the part of the problem nobody is commoditising.

Sources: Hugging Face — DeepSeek-V4.1-Flash-1017 · MarkTechPost — DeepSeek just released V4.1-Flash-1017: scores 77.0 on DeepSWE, opens Search API · AI Weekly — DeepSeek Opens V4.1-Flash-1017 Weights and $0.75 Search API · arXiv — DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression

🌊 Waves

Anthropic's TPU lease is reportedly being structured as $60B of credit, and the junior money wants more than 10%

The junior terms are a report, not a filing: the Financial Times, citing people familiar with the matter, says Blackstone-managed funds have been offered a coupon above 10% to put $9B of their own capital into an $18B Class B junior tranche, syndicating the rest, inside a $60B package arranged by Broadcom and known internally as Project North Star. Above it sits a $42B Class A senior-secured tranche led by banks. The money finances Anthropic's five-year lease of Google TPUs — the $42B covers roughly a third of the $125.2B Anthropic has committed under that lease, with about 3.5 gigawatts of TPU capacity expected to start coming online in 2027. The instruments can convert into Anthropic equity. Two details deserve more attention than the headline number. Comparably rated borrowers pay around 7% in public bond markets, so the junior tier is being priced at a meaningful premium to Anthropic's nominal credit — the market is charging for something the rating does not capture. And Anthropic's own prospectus flags the conflict directly: Broadcom is both the chip supplier and the lender, and certain payment or performance defaults could make a substantial portion of the lease obligations due at once and limit its ability to draw on the $42B facility. Reuters could not immediately verify the report, and the structure has not been established as closed. The deal as described extends the AI XPV partnership Broadcom, Apollo and Blackstone struck in June, which raised $35B against a goal of financing more than 20 gigawatts of AI compute by 2028.

Roadmap implication: This is the compute-financialization wave acquiring a price, and the price is the information. If a lab approaching a $2T public valuation pays double-digit coupons for its compute, then the cost of capacity — not the cost of tokens — is the binding constraint on how fast capacity arrives, and every multi-year rate card you sign sits on top of someone's debt service. Budget for compute contracts to reprice with credit spreads rather than with silicon, write break clauses that survive a lender's bad quarter, and treat supplier-as-lender structures as a single point of failure in your vendor risk register rather than two independent ones.

Sources: AI Weekly — Anthropic Offers Blackstone 10%+ Yield in $9B Chip-Debt Deal · Startup Fortune — Anthropic offers Blackstone a double-digit yield to fund its chip habit · Reuters via AOL — Anthropic offers Blackstone double-digit yield in chip-financing deal, FT reports

Open weights now carry seven in ten tokens through Vercel's gateway, and a quarter of the dollars

Vercel published its October AI Gateway production index on Thursday, covering September traffic, and The Register's read of it followed. Open-weight models handled 71.5% of gateway tokens and took 23% of spend — a share of dollars that has more than quadrupled in three months, from roughly one gateway dollar in twenty in June to nearly one in four in September. The average price per token fell 47% from December while total spend ran about thirteen times December's level, which is the cost-collapse curve and the demand response in one chart. Closed models still own the top of the P&L: Claude Opus 5.5 took 11.03% of spend, GPT-6 Astra 4.86%, DeepSeek V4.1 Flash 4.19%, Claude Sonnet 5.5 3.21%. The routing number is the one to sit with — almost one in three teams changed their most-used lab in September, an index record, with the chart putting lab switching at 28.6% and model switching at 48.0%. And the geography of the money has inverted: more than two-thirds of September gateway spend on Chinese-origin models went to US-based providers serving them, against less than a tenth in December. Moonshot took 3.6% of tokens but 13.2% of spend with nine in ten of its tokens served by US providers; Z.ai 7.2% of tokens and 4.5% of spend, seven in ten served stateside; DeepSeek 58% of tokens and 6.6% of spend, mostly served by DeepSeek itself. Epoch AI's own measurement, published Friday, puts open-weight models roughly six to seven months behind closed ones on its cyber-focused capabilities index, with Z.ai's GLM-5.3 the strongest open entry — so adoption is running ahead of the capability gap rather than waiting for it to close. Vercel's own caveats are load-bearing: spend is estimated at published list prices, the open-weight classification is broader than in earlier editions, and prior months may be revised.

Roadmap implication: This is the measurement the open-model-measurement and model-routing waves have been waiting for, and it says the architecture argument is over in practice. Build on the assumption that your primary model will change inside a quarter — abstract the provider, keep evals portable, and make switching a config change rather than a project. The second opening is the serving layer: if two-thirds of the money spent on Chinese weights is being collected by American hosts, the arbitrage is not in the weights, it is in being the one who serves them under a contract a Western procurement team will sign.

Sources: Vercel — Open weight takes 71.5% of tokens and 23% of spend · The Register — Open-weight models win hearts, minds, and wallets · Epoch AI — Open-weight models trail closed models in cyber capabilities by 6-7 months

The American open-weight field is back within reach, and the licence reads like a term sheet

Nathan Lambert's read on the week, published Saturday, is that the balance of American open-weight models took a step change back toward the frontier — and his framing is worth more than the leaderboard. Meta's Muse Spark Open, 816B total and 42B active with a 1M-token context, is now the strongest American open model by some margin; add Reflection's Beam and, in Lambert's accounting, without those two there would be twenty-plus Chinese open-weight models better than anything the West had downloadable. He puts Muse Spark Open's performance above GLM 5.2, near DeepSeek V4.1 Flash, behind GLM 5.3 and Kimi K3 — the gap between best-American and best-Chinese roughly halved, which moves US open weights from irrelevant-at-the-cutting-edge to behind but usable. The catch is the paperwork. Meta wrote a custom licence carrying nearly every restrictive feature the open-model world has spent three years complaining about: a "Built with Muse" attribution requirement, a revenue gate at $20M per month of total company revenue above which you negotiate separately with Meta, a binding acceptable-use policy, and terms that reach outputs and derivatives — leaving real ambiguity about synthetic data. Internal research and applications are permitted. Lambert's own verdict is that most builders will ignore it and Meta will chase only egregious violations, which is a prediction about enforcement, not a legal opinion. He also notes Mistral has announced Large 4 with an index score of 38 but has not yet released weights, and asks the question the whole field is asking: where have Kimi and Zhipu been since K3 in July and GLM-5.3 in August.

Roadmap implication: The open-weight-credibility wave has stopped being a question about capability and become a question about counsel. If you are standing an open model up in production, the licence is now a selection criterion sitting alongside the benchmarks — run the $20M revenue gate and the derivative-and-output language past legal before you fine-tune, because a model you cannot distil from is a different asset than one you can. The opening for anyone shipping open weights is correspondingly obvious: a plain Apache or MIT licence is now a competitive feature, and Reflection and DeepSeek are both charging into that gap.

Sources: Interconnects — Meta shakes up the frontiers, open and closed · Hugging Face — meta-models/Muse-Spark-Open-816B-A42B · TechCrunch — Meta launches an open-weight AI model that runs on Macs

An AI lab co-founder gave finance ministers the growth forecast

Jack Clark, Anthropic co-founder and head of public benefit, told the IMF and World Bank annual meetings in Bangkok on Saturday that AI's impact is accelerating quickly, and put numbers to it: three internal scenarios for US output in 2030 of $34.1 trillion, $36.3 trillion and $44.4 trillion, against a US economy just over $30 trillion today, with the top case reported as roughly 32% above baseline. He paired the forecast with a call for international regulatory cooperation. Treat the figures as what they are — a vendor's scenario set, delivered by one of the vendor's co-founders, to the audience best placed to act on it, in the same week his company was reported paying double-digit coupons to finance the compute the scenarios depend on. The notable thing is not the number. It is the venue: central bankers and finance ministers do not usually get their productivity assumptions from a model supplier, and a scenario that lands in a finance ministry's planning deck has a way of becoming the baseline it was supposed to be measured against.

Roadmap implication: Macro forecasts shaped by suppliers end up inside the rules those suppliers operate under, which is the governance-as-market-structure tide working through the front door rather than a statute. If your business depends on how a government reads AI's growth contribution — procurement, energy siting, skills funding, tax treatment of compute — this is the number to know and the number to pressure-test, because somebody will quote it at you within the quarter. The opening is to show up with your own evidence: sector-level productivity data from your own deployments is scarce, credible, and currently being written by people with a product to sell.

Sources: Reuters via Investing.com — Anthropic's Clark says AI's impact accelerating quickly and could vastly boost US GDP · Crypto Briefing — Anthropic's Jack Clark says AI could push US GDP to about $44 trillion by 2030 · U.S. News — Anthropic's Clark Says AI's Impact Accelerating Quickly and Could Vastly Boost US GDP

🌊 Ripples

An agent platform lost customer provider keys to people who wanted the models, not the data

OpenCode disclosed on Saturday that it had suffered a breach it describes as a large-scale distillation attack, affecting about 1.4% of Console users who had linked external AI provider accounts. Attackers exploited vulnerabilities in a legacy system to take stored API keys for Anthropic, OpenAI and Google models, then used those keys to run high volumes of queries — harvesting good outputs to train on, at the key holders' expense. The stolen asset was the right to make inference calls, not stored user content; OpenCode has not reported model weights or source code being taken, and says it is working with all three providers on remediation. Affected users are being told to rotate or revoke keys.

Do this now: Rotate every provider key you have ever pasted into a third-party agent platform, today, and then go read the usage dashboards at Anthropic, OpenAI and Google for anomalous spend in the last week — the first sign of this attack is a bill, not an alert. Structurally: stop storing long-lived provider keys in other companies' consoles. Scope keys per integration, cap spend per key, and prefer platforms that broker inference through their own credentials so a breach costs them money rather than you.

Sources: Crypto Briefing — OpenCode discloses distillation attack that hit 1.4% of Console users · Hugging News — OpenCode confirms distillation attack hit 1.4% of Console users

Meta deleted the word "train" from terms that take effect on 30 October

Meta's pending Commercial Terms, due to take effect 30 October, carried a crawling provision letting it use content from business-linked domains to "provide, develop, improve, research, train" its products. The News/Media Alliance objected on Friday, its chief executive Danielle Coffey calling the terms outrageous and probably illegal, and Meta's Andy Stone replied that there had been no change to how Meta uses web crawlers and that this was not about scraping news sites. The text was then revised. The crawling authorisation survives — Meta may still crawl and periodically refresh those domains — but the stated purpose now runs to "advertising and business products" plus integrity and security, and the word "train" is gone from the provision. The licensing section covering content business users share was shortened, and the Self-Serve Ad Terms preview carries the same revised clause.

Do this now: If your company has a Meta business presence, have someone read the 30 October Commercial Terms against the version you reviewed — a clause that changes twice in two days will change again, and the crawling permission itself was never withdrawn. More generally, this is the training-data-liability wave operating pre-litigation: a trade body's objection moved contract language in under 48 hours, which is cheaper and faster than a court. Publishers and platform-dependent businesses should note the price of organised objection just went down.

Sources: Unite.AI — Meta Rewrites Web-Crawling Clause in Pending Commercial Terms · The Register — Meta grants itself right to scrape content from business customers for model training

Anthropic's first Canadian data centre: 95 MW in late 2027, 420 MW by 2028, paid for with wind

An Anthropic executive confirmed on Friday that the company has signed a long-term lease with EverWind Fuels for capacity at a new Point Tupper campus in Nova Scotia — its first data centre agreement in Canada. EverWind says it will invest more than C$10 billion across the campus and associated wind projects, which it calls the largest privately funded investment in the province's history. Phase one is 95 MW targeted for late 2027; a 325 MW second stage would take the site to 420 MW in 2028, subject to commercial, technical and regulatory conditions. Build-out is put at around 2,000 construction jobs and roughly 215 permanent operating positions across campus and wind portfolio. EverWind has committed to fund the electricity-system costs attributable to the project, and Jacob Glick, Anthropic's head of public policy for Canada, said the company would protect ratepayers and water resources and ensure surrounding communities benefit. For scale on why the siting matters: Anthropic spent $7.33 billion on computing and infrastructure last year, roughly triple its 2024 figure.

Do this now: The ratepayer and water commitments arriving with the announcement rather than after the first town-hall is new, and it is the template worth copying — the datacenter-backlash wave has taught the builders to pre-pay the political cost. If you are siting anything power-hungry, the lesson is that a credible answer on who funds grid upgrades is now part of the permit, not a concession you make later. And if you are in Atlantic Canada, 2,000 construction jobs and a 420 MW campus is a supply-chain and services opportunity with a published timeline on it.

Sources: Reuters via The Star — Anthropic signs lease for first Canadian data center · TokenPost — Anthropic Leases 95-Megawatt Phase of Nova Scotia Data Center · Investing.com — Anthropic signs first Canada data centre lease as EverWind plans C$10B investment

Read this edition and the archive at excelsiorgroup.ai/insights/signal