Claude is available two ways for an enterprise buyer: directly from Anthropic, and through Amazon Bedrock as one of the models in the catalog. The per token rates for the same model are broadly comparable across the two, which leads a lot of buyers to assume the choice is a wash and to default to whichever cloud they already use. That assumption costs money, because the rate is the part that is similar. Everything around the rate, how you commit, how you discount, how you negotiate, differs, and the differences are where the real cost lives.
We negotiate Claude deals and study nothing else, so this is a buyer side comparison of the two paths on the terms that actually move a bill. The goal is not to crown one winner for everyone. It is to show you which questions decide it for your situation.
Start with the thing buyers fixate on. For a given Claude model, the headline input and output rates are in the same neighborhood whether you buy direct or through Bedrock. If your whole decision were the sticker rate on a single model, you could flip a coin. But the sticker rate is the least negotiable, least differentiated part of the picture, and treating it as the decision is how buyers end up on the wrong path.
What differs is the packaging around the rate. The commitment mechanics, the discount structure, who you negotiate with, how caching and batch are billed, and how the spend folds into your other agreements, all of these vary by path, and any one of them can outweigh a small rate difference across a real annual workload.
This is the difference that matters most and the one buyers notice least. Buy direct, and you negotiate with Anthropic's account team about a Claude commitment, full stop. Buy through Bedrock, and your leverage is tangled up in your broader cloud relationship and its commitment program, where Claude spend is one line among many.
Each has an edge. The direct path gives you a counterpart whose entire incentive is Claude adoption, which can mean real room to move on a Claude specific deal, and it keeps the negotiation clean and focused. The marketplace path lets Claude spend count toward a larger cloud commitment you may already be making, which can be valuable if you are deep in that ecosystem and want every dollar working toward one threshold. The wrong move is to negotiate neither deliberately and let Claude spend accrete at undiscounted rates on whichever path you drifted onto.
Direct from Anthropic, committed spend earns a discount band, with recognizable thresholds where the rate steps down as the annual commit grows. You negotiate that band, the overage rate, price protection, and how unused commitment is treated, as a Claude specific contract. The terms are the lever, and a negotiated direct deal can pull well ahead of standard pricing on the strength of those terms.
Through Bedrock, Claude consumption typically rides the cloud provider's own commitment and savings constructs rather than a separate Claude band. That can be efficient if you are already committing at scale to that cloud, because the Claude spend helps you reach or clear a threshold you were going to hit anyway. But it also means your Claude discount is mediated by the marketplace mechanics rather than negotiated head on with the people who set Claude pricing, which can leave the model specific room on the table.
The biggest savings on Claude come from how you run the workload, not from where you buy it, and the major levers are available on both paths, though the details differ. Routing each request across Opus, Sonnet, and Haiku to the cheapest model that clears the quality bar typically cuts aggregate spend by forty to seventy percent, and that lever works regardless of path. Prompt caching, which bills repeated context at up to ninety percent off on cached reads, and batch processing at roughly half price for asynchronous work, are both available, but the exact pricing, limits, and feature parity can vary between direct and marketplace.
For a buyer this means two things. First, do not choose a path expecting it to do your optimization for you, because the levers that matter most are yours to pull either way. Second, when feature parity or caching and batch pricing genuinely differs for your specific workload, that difference belongs in the comparison alongside the commit mechanics, not as an afterthought.
Several non price factors quietly decide the path for many buyers, and they are worth weighing honestly. Data residency and the regions each path serves can be the deciding factor for a regulated workload. Existing security reviews and contracts with one provider can lower the friction of using that path. Latency and the proximity of inference to your other infrastructure can matter for interactive products. And procurement simplicity, one more line on an existing agreement versus a new vendor contract, has a real cost in time even when it does not show up on the invoice.
None of these is about the rate, and all of them can outweigh it. The mistake is to let one of them, usually procurement convenience, silently decide the path without checking what it costs on the commercial terms.
The decision comes down to a few honest questions. How large is your Claude spend relative to your total commitment to a given cloud, and would routing it through that cloud help you clear a threshold you were going to commit to anyway. How much room is there in a direct, Claude specific negotiation given your volume. Do data residency, latency, or compliance constraints favor one path on their own. And which path lets you actually negotiate rather than accept standard pricing.
For many large, Claude heavy buyers, a direct Anthropic deal negotiated properly wins on terms, because the band, the overage rate, and the protections are all on the table with the people who own Claude pricing. For buyers already committed deeply to a cloud where Claude is a modest slice, the marketplace path can be the efficient choice. The answer is specific to your numbers, which is exactly why it should be worked rather than assumed.
The download below walks through the comparison with the current commit bands and the terms worth negotiating on each path, and you can also see our full read on Anthropic pricing in 2026.
Download the buyer side comparison of direct and marketplace Claude pricing, with the commit bands and terms that decide each one.
Download the playbookWeekly intelligence on Anthropic pricing moves and the buyer side counters that work.