THE AI ARBITRAGE

Frontier Intelligence, Half the Price

Anthropic's third flagship release in five weeks undercuts its own top model on cost while claiming the lead on several of its internal benchmarks. For any operator routing work to Claude by default, this is the week to check which model is actually running.

Claude Opus 5 went live Friday, July 24, priced at $5 per million input tokens and $25 per million output tokens, the same rate as its predecessor Opus 4.8 and exactly half of what Claude Fable 5 costs. Anthropic's own announcement frames it plainly: Opus 5 "comes close to the frontier intelligence of Claude Fable 5 at half the price." On coding and knowledge-work evaluations the company runs internally, including Frontier-Bench v0.1, Anthropic says Opus 5 is now the strongest model in its lineup, though it remains behind Anthropic's own Mythos 5 on cybersecurity-specific tasks by design.

Anthropic is positioning Opus 5 less as a bigger model and more as a more efficient one. On CursorBench 3.2 at maximum effort, the company says Opus 5 scores within half a percentage point of Fable 5's peak result at half the cost per task, and on OSWorld 2.0, a computer-use benchmark, it says Opus 5 beats Fable 5's best score at just over a third of the cost. Opus 5 is now the default model on Claude Max and the top-ranked option on Claude Pro.

Anthropic also says Opus 5 is its most aligned model to date on internal behavioral audits, and that its cybersecurity safety classifiers intervene roughly 85% less often than they do for Fable 5, meaning fewer requests get blocked or silently downgraded to a fallback model. Those are Anthropic's own claims about its own model, worth treating as a vendor's self-assessment rather than an independently audited result, but the pricing and default-model changes are independently verifiable and already in effect.

The read for operators: if a workflow is routed to Claude by default rather than by a deliberate choice of model, this is the week that default gets cheaper without anyone having to change anything, or it stays needlessly expensive if the routing was never revisited. See The six50 POV below.

The Bill Nobody's Itemized Yet

Enterprise AI spending keeps climbing while per-token prices keep falling. CIOs still can't reliably say what a single workflow costs. For a $2M-$50M operator, that gap is where budgets go to die quietly.

"Like with every scalable enterprise technology component, the bill for massive AI deployments eventually comes due," CIO Dive wrote in a Friday, July 24 roundup of its recent enterprise AI cost coverage. The piece's own framing is the story: with the experimentation phase largely behind them, technology leaders have spent much of 2026 trying to bring runaway AI spending under control as adoption spreads through their organizations, and it's "a tricky balance," since too tight a rein blocks innovation while unclear guardrails and application sprawl wreck budgets right as leadership expects returns. Vendors have started responding, CIO Dive noted, with Oracle and AWS rolling out new billing visibility tools and the Linux Foundation launching a dedicated token-cost-management initiative earlier this summer.

The same week that roundup published, three frontier vendors put three different per-million-token prices into the market for their flagship-tier models: Anthropic's Opus 5 at $5 input / $25 output, Anthropic's own higher-tier Fable 5 at $10/$50, and Moonshot AI's Kimi K3 at roughly $3/$15 (see the chart in the lead story above). None of these is "the" price of AI. Each is a bet on how much model a given task actually needs, and for most day-to-day SMB workflows, the honest answer is less than the top-tier default.

The read for operators: this is precisely the blind spot the First 90 Days Diagnostic's AI automation roadmap is built to catch. A workflow that looks automated and cheap on the surface can be one routing decision away from a very different monthly number, and most SMB operators won't notice until the invoice lands.

Beijing's Answer Arrives, Weights and All

Days after a White House official accused Moonshot AI of copying its way to a leaderboard finish, the company is making its case in public: full open weights, the largest release of its kind, free to download.

Moonshot AI's Kimi K3, a 2.8-trillion-parameter mixture-of-experts model that activates only 16 of its 896 experts per token, is releasing its full open weights on Hugging Face this weekend, at 00:00 UTC July 27, roughly 8 PM Eastern on July 26. At an estimated 594 gigabytes using MXFP4 quantization, multiple outlets covering the release describe it as the largest open-weight model release to date. Kimi K3's API has already been live since mid-July at roughly $3 per million input tokens and $15 per million output tokens, undercutting Anthropic's Opus 5 by close to half.

The timing is hard to miss. Our July 24 issue covered a top White House science and technology official's accusation that Moonshot covertly distilled Anthropic's Fable model at industrial scale to build K3. Moonshot hasn't directly rebutted that claim. Publishing the full weights is, in effect, an answer of its own: anyone can now inspect, fine-tune, or self-host the model rather than take either side's benchmark claims on faith. Third-party leaderboard trackers have reported Kimi K3 near the top of some coding benchmarks, ahead of several closed models; we're not reproducing specific benchmark rankings here because they come from a single tracking source we could not independently corroborate before this issue's deadline.

The read for operators: the parameter count is not the number that matters for a $2M-$50M business, the 594 GB download puts self-hosting out of reach for nearly all of them. The price line does matter: a credible frontier-adjacent model at $3/$15 puts real pressure on every closed-model vendor's next pricing move, Anthropic's Opus 5 cut included.