The prices published for Claude on the cloud marketplaces match Anthropic's own list prices: as of July 2026, $5.00 per million input tokens for Opus 4.8, $3.00 for Sonnet 5 (with an introductory $2.00 rate through August 31, 2026), and $1.00 for Haiku 4.5, with output priced higher in each tier. For a pilot, list price is what you pay and that is fine. Once usage becomes predictable, though, list price is a starting point, and the discount conversation runs through your cloud provider, not Anthropic.
Why the discount comes from the cloud, not the model maker
When you consume Claude on a third-party platform, your commercial relationship is with AWS, Google Cloud, or Microsoft. They bill you, their marketplace terms apply, and consequently they are the party with the authority to discount. Anthropic sets list prices; committed-use arrangements are negotiated with the cloud provider. This has three practical consequences. First, your account team is the entry point, the same people who handle your existing enterprise agreement. Second, Claude spend can often count toward broader cloud commitments you have already signed, which matters if you are working to burn down a multi-year spend obligation. Third, the shape of the deal (term length, drawdown rules, what happens if you underconsume) follows your cloud's commercial norms rather than a separate AI-specific contract.
What a commitment actually is
A committed-use arrangement is a simple trade: you promise a level of spend or consumption over a defined term, and the provider gives you better-than-list pricing in return. The specific discount rates, thresholds, and terms vary by provider, by deal size, and over time, so treat any number you hear second-hand as stale. What is consistent is the structure of the risk: if you overcommit and underconsume, you pay for capacity you did not use; if you undercommit, you simply pay list price for the overage, which is a much gentler failure mode.
Preparing for the negotiation
Walk in with numbers, because the discount you get tracks the credibility of your forecast. Useful preparation looks like this:
Bring token-level usage history broken out by model tier, since input, output, and tier mix determine spend more than request counts do. Bring a growth forecast tied to concrete launches ("feature X ships to all customers in Q4") rather than aspiration. Know your optimization roadmap: if you plan to adopt prompt caching or route more traffic to Haiku 4.5, your future spend may be lower than a straight-line projection, and committing against the unoptimized number locks in waste. Finally, ask explicitly how Claude consumption interacts with your existing enterprise discount program and marketplace commitments; the answer differs by provider and can change which platform is cheapest for you in practice.
Traps worth knowing about
Three patterns cause regret. The first is committing during a pilot, when usage is unrepresentative and prompts are unoptimized; the optimization work you do in the first production quarter routinely changes spend enough to make an early commitment look foolish. The second is committing to a single model tier when your tiering strategy is still evolving; where possible, prefer commitments denominated in spend rather than in consumption of one specific model. The third is letting a commitment silently discourage good engineering: if the team knows the company must burn a committed amount, the incentive to cache prompts and tier models evaporates. Keep the FinOps discipline running regardless of what was signed.
Also confirm the mundane details in writing: which regions and models the discounted rate covers, how introductory prices (like Sonnet 5's current one) interact with the discount, and how the deal treats new models released mid-term. None of this is exotic, but all of it is easier to fix before signature than after.
Where to go next
Start with Claude 3P Pricing, Explained Like a Utility Bill if the token-based billing model is new to you, and see Working with Procurement and Legal for the contracting side of marketplace purchases. The platforms overview summarizes who operates what.