Claude Sonnet 5's $2 per million input tokens does not undercut GPT-5.6 Terra's $2. Anthropic's own documentation states that Claude 4.7 and later models use a newer tokenizer that "produces approximately 30% more tokens for the same text" — so on an identical workload, Sonnet 5 bills about 15% more than Terra today, and about 73% more from 1 September when its introductory rate expires. Normalise on text, not tokens, or the rate card will pick your vendor for you.
The Verdict, and the Workload It Assumes
Run high-volume production traffic on Gemini 3.6 Flash, escalate the genuinely hard reasoning to Claude Opus 5, and keep GPT-5.6 Terra as the second provider behind a gateway. Do not standardise on one. Every number below moves faster than your procurement cycle.
The products under comparison are the API tiers of the three frontier families — Claude, Gemini and the models behind ChatGPT — not the seat-priced end-user suites, which are a different purchase with different economics.
Everything here is normalised to one workload, stated so you can argue with it: a document-grounded support-triage agent at 300,000 requests per month (10,000/day), 8,000 input tokens per request of which 7,000 are a stable cached prefix, and 600 output tokens. Token counts are measured on GPT-5.6's tokenizer, then adjusted for Anthropic's, which is the entire point of this article. Prices checked 5 August 2026 against each vendor's live pricing page.
| Model | Rate card (in / out per MTok) | Monthly cost, this workload | Output speed | Time to first token | Buy it for |
|---|---|---|---|---|---|
| GPT-5.6 Luna | $0.20 / $1.20 | $318 | — | — | Classification, extraction, routing |
| Gemini 3.6 Flash | $1.50 / $7.50 | $2,120 | 212.7 tok/s | 15.7s | Default production volume |
| GPT-5.6 Terra | $2.00 / $12.00 | $3,180 | — | — | Second provider, broadest tooling |
| Gemini 3.1 Pro Preview | $2.00 / $12.00 | $3,203 | 128.7 tok/s | 31.5s | Nothing. See below. |
| Claude Sonnet 5 (to 31 Aug) | $2.00 / $10.00 | $3,666 | 82 tok/s | 200.6s | Agentic work, 1M context |
| Claude Sonnet 5 (from 1 Sep) | $3.00 / $15.00 | $5,499 | 82 tok/s | 200.6s | The same work, 50% dearer |
| GPT-5.6 Sol | $5.00 / $30.00 | $7,950 | 69 tok/s | 143.5s | Hard reasoning, OpenAI stack |
| Claude Opus 5 | $5.00 / $25.00 | $9,165 | 55.1 tok/s | 52.5s | The hardest reasoning |
Rate cards from Anthropic's pricing docs, OpenAI's API pricing and the Gemini API pricing page. Speed and latency from Artificial Analysis.
Read the two columns against each other. Opus 5 lists a cheaper output rate than GPT-5.6 Sol — $25 against $30 — and costs 15% more to run. Sonnet 5 lists cheaper than Terra and costs 15% more. The rate card inverts the answer in both cases.
Why the Per-Token Price List Misleads by 30%
A tokenizer is the function that chops your text into the billable units a model charges for. It is model-specific, and Anthropic changed theirs.
The pricing documentation carries the warning in a footnote: Claude 4.7 and later models "use a newer tokenizer that contributes to their improved performance" and it "produces approximately 30% more tokens for the same text. The exact increase depends on the content and workload shape." Claude Sonnet 4.6 and earlier use the previous tokenizer. Sonnet 5 and Opus 5 both use the new one.
That single sentence is worth more than every benchmark table published this quarter, because it means a price comparison denominated in tokens is not a comparison at all. The same support ticket, the same contract, the same 40-page policy PDF becomes a bigger bill on Claude than the rate card implies — not because Anthropic charges more per token, but because your document is more tokens.
The magnitude is workload-dependent, and the independent measurements land close to Anthropic's number — sometimes past it. OpenRouter analysed over a million requests from users who switched across the tokenizer change and found 32–45% raw inflation, with the effective billing increase at 27.2% for prompts in the 2,000–10,000 token band — which is where the workload in this article sits. It fades at both ends: under 2,000 tokens it went slightly negative, because the newer models answer short questions more concisely, and above 50,000 tokens prompt caching absorbs most of the extra.
The direction of it is the opposite of what most buyers guess. Measured through Anthropic's own token-counting endpoint, inflation is heaviest on English technical documentation (1.47x) and prose-like config files (1.44x), and lightest on exactly the content people expect to be worst hit — dense JSON (1.13x), numeric CSV (1.07x), and Japanese and Chinese prose (1.01x, effectively unchanged). Simon Willison separately measured 1.46x on text and 3.01x on images. If your corpus is structured data or CJK text, your ratio is near 1.0 and most of this section does not apply to you.
The honest way to settle it costs you an afternoon. Take 500 real production requests, run them through Anthropic's token-counting endpoint and OpenAI's, and compare the counts on your corpus. Every number in this article is a prior. Your own token ratio is evidence.
What One Real Workload Actually Costs Per Month
Working the arithmetic on the workload above, at 2,100 MTok of cached reads, 300 MTok of fresh input and 180 MTok of output per month:
- GPT-5.6 Terra: 2,100 × $0.20 + 300 × $2.00 + 180 × $12.00 = $3,180.
- Claude Sonnet 5, same text, 30% more tokens: 2,730 × $0.20 + 390 × $2.00 + 234 × $10.00 = $3,666.
Before the tokenizer adjustment Sonnet 5 looks like $2,820 — 11% below Terra. After it, Sonnet 5 is 15% above. On 1 September the introductory rate lapses to $3/$15 and the same traffic becomes $5,499, or 73% above Terra, on a workload that has not changed by a single request.
That 15% is contingent on the 1.30x ratio, so here is the number it turns on: this workload breaks even at 1.13x. Below that Sonnet 5 is genuinely cheaper than Terra, which is why the CSV and CJK workloads above should stop reading here and go measure. The September comparison is not contingent on anything — at $3/$15 the same traffic costs $4,230 even if your text tokenized identically on both, a third more than Terra before the tokenizer is mentioned at all.
The same pricing page states the transition plainly: $2/$10 through 31 August 2026, $3/$15 from 1 September. If you are budgeting off a Sonnet 5 pilot run this month, your Q4 forecast is 50% light before you account for anything else.
Two mechanics worth knowing before you build the model:
Prompt caching is the largest lever any of them offer, and they price it differently. Anthropic charges a write premium (1.25x base for a 5-minute TTL, 2x for an hour) and reads at 0.1x. OpenAI's cached input is a flat 10% of base. Google charges reads at 10% plus an hourly storage fee — $1.00 per million tokens per hour on Flash, $4.50 on 3.1 Pro. At a 7,000-token prefix that storage line is about $5 a month and irrelevant. Hold a 500,000-token corpus warm on 3.1 Pro and it is roughly $1,640 a month whether or not you send a request. Anthropic and OpenAI charge nothing for an idle cache. If you are building long-context agents, that asymmetry is the line item nobody models.
Batch is 50% off at all three and almost nobody uses it. If a workload can tolerate asynchronous completion — nightly enrichment, backfills, evaluation runs — it halves. Note that on Anthropic the Batch API carries 29-day retention and sits outside a zero-retention arrangement, which matters if you signed one.
Latency Is Now the Binding Constraint, Not Price
The reasoning models have become slow in a way that removes them from entire categories of work, and money only partly fixes it.
Measured by Artificial Analysis, GPT-5.6 Sol at maximum effort has a time to first token of 143.53 seconds, against a median of about 2.80 seconds for comparable models. Claude Opus 5 is 52.47 seconds. Claude Sonnet 5 at max effort is 200.59 seconds. Gemini 3.6 Flash is 15.67 seconds at high effort and 212.7 output tokens per second — an order of magnitude ahead of the field on both.
These are configured-at-maximum-effort numbers, and effort is a dial you control on all three. But the shape of the finding survives the caveat: if a human is waiting on the response, the frontier reasoning tier is not a candidate, and the gap between Flash-class and reasoning-class latency is now wider than the gap between their intelligence scores.
You can pay to fix half of this, and it is worth knowing which half. Anthropic sells fast mode — research preview, access through your account manager — at $10/$50 per MTok, double the standard Opus 5 rate, for up to 2.5x higher output tokens per second. Read its own documentation before you budget for it: the benefit is "focused on output tokens per second (OTPS), not time to first token (TTFT)." It makes a long answer finish sooner once it has started. It does not shorten the 52 seconds of silence before that. It is also unavailable on Bedrock, Google Cloud, Microsoft Foundry, the Batch API, and alongside a Priority Tier commitment. Opus 5 leads the Artificial Analysis intelligence index at 61 against Sol's 59 and Sonnet 5's 53 — an eight-point spread across a tier whose time to first token ranges from 52 to 200 seconds.
The practical consequence is architectural, not commercial. Interactive surfaces get Flash-class models with a hard timeout. Reasoning-class models go behind a queue, a job status endpoint and a progress indicator — the same pattern you would use for a report build. Teams that skipped that step and wired Opus straight into a chat box are the ones filing latency tickets against a model that is behaving exactly as documented.
Where Your Data Lives Costs the Same 10% Everywhere
Three vendors, three independent pricing teams, one number. If you are regulated, your real rate card is 10% above the published one.
- Anthropic:
inference_geo: "us"applies a 1.1x multiplier across input, output, cache writes and cache reads on Claude 4.6 and later. Global routing is the default and is standard price. - OpenAI: data residency spans regions including the EU, UK, Japan, India, Canada and Australia, and carries a "10% uplift" for eligible models released on or after 5 March 2026.
- Google: pricing varies by endpoint and is generally lower on global endpoints; regional and multi-region endpoints carry a 10% premium on Claude models. Only regional endpoints guarantee in-region processing.
The pricing convergence is tidy. The capability divergence is not, and it is the thing that should actually move a European buyer.
Anthropic's first-party data residency documentation states that workspace geo — which governs where data is stored at rest — is currently "us" and only "us", and cannot be changed after a workspace is created. Inference geo offers exactly two values, "us" and "global". There is no EU option. An EU-resident Claude deployment therefore has to run through Amazon Bedrock, Google Cloud or Microsoft Foundry — and on the first two, per Anthropic's own retention docs, "the cloud provider is the data processor" and Anthropic's zero-retention and HIPAA arrangements do not apply. You get the cloud's controls instead, which are good, but they are not the ones your legal team read.
OpenAI offers per-project EU residency with zero data retention directly. Google offers at-rest residency across a long list of countries. On this axis the ranking inverts the intelligence ranking completely, and for a bank in Frankfurt that is the only ranking that matters.
Two more retention details that get missed:
- Zero data retention is sales-gated at both OpenAI and Anthropic, not a setting. Neither is self-serve.
- ZDR does not cover everything you will actually use. On Anthropic it excludes the Batch API (29 days), the Files API, code execution containers (30 days), Managed Agents and the MCP connector. On OpenAI,
/v1/assistants,/v1/threadsand/v1/vector_storesare not eligible. And Claude Fable 5 and Mythos 5 — Anthropic's top-priced tier at $10/$50 per MTok — require 30-day retention and cannot be run under ZDR. There is a documented escape hatch that is easy to miss: a ZDR organisation can enable 30-day retention on one workspace and keep zero retention across the others. That is a deliberate, auditable hole in your own compliance story rather than a checkbox, but it is not the flat exclusion it first reads as.
Who Can Actually Sell You a Capacity Guarantee
This is the row that changed most in 2026, and it is buried in documentation nobody reads before signing.
Anthropic's service tiers page now opens with a warning: "Priority Tier capacity commitments are no longer available for purchase." Existing commitments run to their contract end date. And the supported-model list excludes Claude Opus 5 and Claude Sonnet 5 outright — the two models you would actually deploy. Guaranteed capacity on Anthropic's current flagship is a sales conversation, not a product.
OpenAI still sells Scale Tier as purchasable input and output token units, bought for a minimum of 30 days, with a 99.9% uptime SLA attached. Google sells Provisioned Throughput in generative AI scale units on terms from a week upward, and has extended it across Gemini, open-weight models and — in private preview — Claude.
If your board has asked for an availability commitment in writing, that is the entire analysis. Two vendors will sell you one off the shelf. One will not.
The Indemnity Has a Hole Shaped Like Your Use Case
All three indemnify. None indemnify what you probably think.
Anthropic's Commercial Terms assign output ownership to the customer and commit Anthropic to defend third-party IP claims arising from paid use in accordance with the terms. Then the exclusions: claims arising from customer modifications, combination with non-Anthropic technology, customer inputs, "practice of a patented invention contained in an Output", and "alleged violation of trademark based on use of an Output in trade or commerce."
Read those last two again against what you are building. Generated code that practises a patent: excluded. Generated marketing copy that trips a trademark: excluded. Those are not edge cases — they are the two highest-volume enterprise use cases in the building.
Google's indemnity is unusual in covering two things: its own use of training data, and generated output. But it applies only to "generally available versions of foundation models" — pre-GA offerings are excluded from it entirely. OpenAI's Copyright Shield extends to API and Enterprise output, but only to generally available, as opposed to client-tailored, features — so confirm in writing that a fine-tuned or customised deployment is still covered before you rely on it.
The action item is one line in your evaluation rubric: name the specific output type your business ships, and find it in the exclusions before you sign.
Gemini 3.1 Pro Preview Is the One to Skip
A comparison where everything wins its own category is marketing. This one loses.
Gemini 3.1 Pro Preview costs $2/$12 per MTok, stepping up to $4/$18 above 200,000 input tokens — input doubles, output rises by half — a cliff exactly where long-context agent work lives. It runs at 128.7 output tokens per second with a 31.53-second time to first token. Google's own Gemini 3.6 Flash costs $1.50/$7.50 flat with no long-context tier, runs 1.6x faster, and reaches first token in roughly half the time.
Then the part that should end the conversation: it is a preview model, per Google's own model list. Google's generative AI indemnity covers generally available foundation models. A preview model is not one. You would be paying 51% more on this workload, waiting twice as long, and standing outside the indemnity — to use the older model from the same vendor.
If you need Google's top reasoning tier today, the honest answer is that Flash has closed enough of the gap that Pro Preview is not worth the exposure. Revisit when Google's next Pro model reaches GA — it is still in partner testing.
Who Should Not Buy Each of These
The most useful column in any comparison, and the one sponsored content never writes.
Do not pick Claude Opus 5 or Sonnet 5 if: you need a contractual capacity guarantee (Priority Tier is withdrawn and excludes both); you need EU data residency on the first-party API (workspace geo is US-only); you have committed to zero data retention and want the Fable or Mythos tier (they require 30-day retention); or your cost model was built on token counts from Sonnet 4.6 (add ~30%). Also not if a human is waiting: 52 to 200 seconds to first token.
Do not pick GPT-5.6 Sol if: your workload is output-heavy. At $30/MTok output it is the most expensive option in this comparison on a real workload, and at 143.5 seconds to first token it is also among the slowest. Terra does most of the same work at 40% of the cost. Sol earns its price on genuinely hard reasoning and nowhere else. If you plan to fine-tune, confirm Copyright Shield still covers the result before you rely on it.
Do not pick Gemini 3.6 Flash if: you need the top of the reasoning distribution — Opus 5 and Sol both score meaningfully higher on the Artificial Analysis intelligence index. Do not pick it if your architecture holds a very large context cached continuously; Google's hourly storage fee makes an idle cache a standing charge that neither competitor levies.
Do not pick GPT-5.6 Luna for anything a customer reads unreviewed. At $318/month for this workload it is nearly seven times cheaper than Flash, and it is a small model. Use it for classification, routing, extraction and first-pass triage, behind something that checks its work.
How to Decide, and What Changes the Answer
Four criteria predict regret. The rest are noise.
This week. Run 500 real production requests through both tokenizers and get your own ratio on paper. Every cost model in your organisation that compares vendors per-token is wrong by an unknown amount until you do.
This month. Put a gateway between your application and the providers, even if you only wire one up. Two of the four variables here — introductory pricing expiring, capacity commitments being withdrawn — moved inside the last 90 days. The cost of the abstraction is a week. The cost of not having it is a migration.
Before your next renewal. Take the specific output your business ships to customers and find it in each indemnity's exclusion list. If it is there, you are self-insuring and should know that before the renewal, not after the letter.
Before you commit spend. Ask each vendor, in writing, what capacity you can contract for on the model you actually intend to deploy — not the model family. The answer differs by vendor today in a way it did not six months ago.
What would change this recommendation: Anthropic opening a non-US workspace geo, which would remove the single largest reason an EU enterprise routes Claude through a cloud provider. Google's next Pro model reaching GA, which would give it an indemnified top-tier model. Or OpenAI shipping a Flash-class model at Gemini's latency, which would collapse the three-provider argument into two.
The Bottom Line
Every enterprise that set out to standardise on one frontier provider in 2026 ended up with two or three, and the reason is not indecision. It is that the four variables that matter — effective cost, latency, where data may sit, and what is contractually guaranteed — do not rank the same vendor first, and they each move on a different clock. Anthropic leads on reasoning and is the hardest to get a capacity commitment from. Google leads on latency and price and puts a meter on idle cache. OpenAI leads on breadth and residency coverage and charges most for output.
Pick a gateway, not a winner. The gateway is the only decision on this page with a shelf life longer than a quarter.
Continue Reading
- Pinecone vs Weaviate vs pgvector: Stay on Postgres
- LangGraph vs CrewAI vs AutoGen: One of Them Is Retired
- Snowflake Cortex vs Databricks Mosaic AI: Pick on Exit Cost
- Your AI Router Is Trading a 10x Discount for a 2.5x One
- GPT-5.6 Is Live: 80% Price Cut Resets Enterprise AI Math
- DeepSeek Swapped the Model. Your Eval Didn't Notice.
- ChatGPT Enterprise vs Claude Enterprise: The $200K Decision
- EU AI Transparency Law Starts Monday: Are You Ready?
