Three of the four newest frontier APIs now share one number: $0.10 per million cached input tokens. OpenAI lists it for GPT-6.1 Sol, Anthropic cut Claude Sonnet 5.5 to it on October 7, and Google’s 95% cache discount on Gemini 4 Argon’s $2 input rate lands on the same figure. For anyone weighing Gemini vs ChatGPT vs Claude on API cost, that convergence matters more than any headline price, because agent workloads spend most of their input tokens on repeated context.
The picture gets less tidy after November 30. Google says Argon’s introductory $2 input and $10 output rates end that day, and the public price becomes $4 and $20 from December 1. Anthropic, for its part, logged Sonnet 5.5’s September 28 launch in its release notes, and the October 7 cache-read cut halved a line item that Anthropic says lowers typical agentic-work costs by about 20%. We covered the Google side separately in our look at Gemini 4 Argon’s price change. This article widens the lens to four vendors and asks what each API actually costs to run.
We include DeepSeek V4.1-Flash as the fourth contender because it sits in a different price class altogether. It is not a like-for-like rival to the other three, but any team that budgets by the token needs to see how far apart the tiers sit. Everything below reflects public vendor material as of October 22, 2026. Where our sources disagree or rest on a single page, we say so in the text instead of smoothing it over.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
Gemini vs ChatGPT vs Claude API pricing on October 22, 2026
Start with the rate card. The table below lists the prices we could confirm against vendor pages fetched today. Some cells say “see note” because the base Sonnet 5.5 rates and the DeepSeek rates rest on a single page each, so they appear in the notes under the table with the hedge they deserve. For the budget tier, our head-to-head of Sonnet 5.5 against GPT-6 Luna covers the cheaper siblings in more depth.
| Model | Input per 1M tokens | Cached input per 1M | Output per 1M tokens | Status on Oct. 22, 2026 |
|---|---|---|---|---|
| OpenAI GPT-6.1 Sol | $2.00 | $0.10 | $10.00 | Official pricing page, updated Oct. 21 |
| Google Gemini 4 Argon (to Nov. 30) | $2 | 95% off input = $0.10 | $10 | Introductory price, public access |
| Google Gemini 4 Argon (from Dec. 1) | $4 | Not restated for the new rate | $20 | Announced increase |
| Anthropic Claude Sonnet 5.5 | See note 1 | $0.10 read ($0.20 before Oct. 7), $2.50 write | See note 1 | Generally available since Sept. 28 |
| DeepSeek V4.1-Flash | See note 2 | See note 2 | See note 2 | Peak and off-peak schedule, off-peak is 50% of peak |
Note 1. Third-party pages list Sonnet 5.5 at $2 input and $10 output per million tokens, and the cache numbers above are consistent with a $2 base (a $2.50 write is a 25% premium on $2, and a $0.10 read is 5% of it). Anthropic’s own pricing page did not expose the base table to our crawler, and several third-party pages disagree with each other, so confirm the figures on Anthropic’s page before you sign a budget.
Note 2. One page reports DeepSeek’s peak rates for V4.1-Flash as $0.30 per million uncached input tokens, $0.006 per million cached input tokens, and $1.20 per million output tokens. DeepSeek’s own pricing documentation confirms the structure (off-peak rates are 50% of peak, effective since September 10), but other pages list different numbers for earlier Flash checkpoints. Treat the figures as indicative.
Price-tracking sites help here. The Tokely tracker of official prices across 66 models is one place to cross-check list rates, though it is a third-party compilation and not a substitute for the vendor page. What stands out in the confirmed numbers is how little separates the three Western vendors. GPT-6.1 Sol and Gemini 4 Argon match on input and output through November, and Sonnet 5.5 reportedly sits in the same band. On list price the spread is zero, which pushes the buying decision toward caching rules, batch discounts, and what happens when promotional rates expire.
OpenAI also published a three-tier text ladder on its pricing page. The tiers run $10 input and $50 output, $2 input and $10 output, and $0.10 input and $0.50 output per million tokens. They line up with GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna in OpenAI’s own advice to start with Astra for complex reasoning and coding, pick Sol to balance intelligence and cost, and use Luna for cost-sensitive, high-volume work. Sol is the mid-tier, and that is the model we compare here.
Specs side by side: release dates, access and discounts
The specs table covers what the vendors have published. Where a cell says “not stated,” our sources did not give a figure, and we chose not to guess. DeepSeek’s own V4.1-Flash announcement is the primary source for the right-hand column.
| Spec | Claude Sonnet 5.5 | GPT-6.1 Sol | Gemini 4 Argon | DeepSeek V4.1-Flash |
|---|---|---|---|---|
| Maker | Anthropic | OpenAI | DeepSeek | |
| Release | Sept. 28, 2026 | Sept. 29, 2026 | Generally available Oct. 13, 2026 | Sept. 10, 2026 (Flash-1017 checkpoint Oct. 17) |
| Family position | Middle of Claude 5.5 (Opus 5.5, Sonnet 5.5, Haiku 5.5) | Middle of GPT-6 (Astra above, Luna below) | Google’s newest Gemini release | Flash tier, single stable alias |
| Context window | 1M tokens, per third-party reports | Not stated in our sources | Not restated here (see Argon section) | 1M tokens, reported |
| Multimodal input | Not stated | Not stated | Not stated | Native multimodal support |
| Input price (list) | See note 1 | $2.00 / 1M | $2 / 1M, then $4 from Dec. 1 | See note 2 |
| Output price (list) | See note 1 | $10.00 / 1M | $10 / 1M, then $20 from Dec. 1 | See note 2 |
| Cached input | $0.10 read, $2.50 write | $0.10 / 1M | 95% below input rate | Separate cached rate, see note 2 |
| Batch discount | 50% (Anthropic pricing page) | 50% (OpenAI pricing page) | Not stated | Not stated |
| Data residency / region surcharge | Not confirmed | +10% data residency | Not stated | Not stated |
| Where you can buy it | Claude Platform, AWS, Google Cloud, Microsoft Foundry | OpenAI API, plus Google Cloud Model Garden (private preview, Oct. 21-22) | Gemini API, AI Studio, Antigravity, AI Mode | DeepSeek API, web and mobile apps |
| Retrieval add-ons | Not stated | Web search priced per 1,000 calls | Not stated | Search and Fetch at $0.75 per 1,000 requests from Nov. 1 |
| Pricing change on the calendar | Cache read cut Oct. 7 | GPT-6 Sol and Luna price cuts announced Oct. 21 | Doubles after Nov. 30 | Search and Fetch billing starts Nov. 1 |
Three details deserve a closer read. First, the Claude Sonnet 5.5 context figure comes from third-party pages, which also report no long-context surcharge on the later Claude models. Anthropic’s own materials in our crawl did not restate it. Second, OpenAI’s pricing page lists a flat web-search charge of $10 per 1,000 calls, so tool costs sit outside token math. Third, DeepSeek’s API alias now points at the latest Flash checkpoint, which makes upgrades automatic and reproducibility harder.
The release calendar is also unusually compressed. Within five weeks, Anthropic shipped Sonnet 5.5 and Haiku 5.5 (October 7), OpenAI shipped GPT-6.1 Sol, Google took Argon from a restricted cyber-defense program to general availability, and DeepSeek pushed a Flash update on October 17. Any benchmark table older than three weeks is probably describing last month’s checkpoint of one of these models.
The $0.10 cache rate: how three vendors landed on one number
Prompt caching stores a long, repeated prefix (a system prompt, a codebase map, a tool schema) so later requests pay a fraction of the normal input price. For agents that re-send the same context on every step, caching is the single biggest lever on the input side of the bill.
OpenAI lists GPT-6.1 Sol cached input at $0.10 per million tokens against $2.00 for fresh input, a 95% discount. Google’s Argon launch terms describe cached input as priced 95% below the input rate, which at $2 also works out to $0.10. Anthropic reached the same place on October 7, when it cut Sonnet 5.5 cache reads by 50%, from $0.20 to $0.10 per million tokens. At the reported $2 base, that move turned a 90% discount into a 95% discount and matched its two largest rivals.
The rates are equal, but the mechanics are not. Anthropic charges a write premium. Sonnet 5.5 cache writes are listed at $2.50 per million tokens, which is 25% above the reported $2 input rate. OpenAI and Google, in the material we reviewed, do not describe an explicit write surcharge. That difference rarely matters for a prefix reused hundreds of times, and it can matter a lot for a prefix reused twice.
A worked cache example
Take a 50,000-token prefix that an agent reuses across 1,000 requests. On Claude, you pay one cache write (50,000 tokens at $2.50 per million, or $0.125), and every later request reads the prefix at $0.10 per million tokens. Without caching, at the reported $2 rate, every request pays the full input price for the whole prefix. The cache cuts that part of the bill by roughly 95%.
Now drop the reuse count to two. The write costs $0.125 and the single read costs $0.005, against $0.20 uncached. The saving shrinks from 95% to about a third, and it vanishes entirely if the cache expires between calls. Read each vendor’s cache time-to-live rules before you assume the headline discount applies to your traffic pattern.
The takeaway: when three vendors quote the same cached rate, the winner on your invoice is the one whose cache behavior fits your request pattern, and not the one with the smallest number on the page.
Gemini 4 Argon’s November 30 price cliff
Google announced Argon on October 1 to a limited group of trusted cyber defenders through its Fairwind Program, then opened it to everyone on October 13 through the Gemini API, Google AI Studio and the Gemini platform. The consumer side followed in the Gemini app and AI Mode in Search for paid subscribers, with Google AI Ultra subscribers also getting Google Antigravity. Our direct comparison of GPT-6.1 Sol against Gemini 4 Argon goes further on the capability side.
The pricing text is the part that changes budgets. Google’s introductory rates are $2 per million input tokens and $10 per million output tokens, with cached input at 95% off. The official page then says that after the introductory period expires on November 30, 2026, the price of $4 per million input tokens and $20 per million output tokens applies. The page adds that this applies to the Argon model intended for public access only.
That is an exact doubling on both axes. The arithmetic is simple, and the consequences are easy to miss because the intro rate matches GPT-6.1 Sol so cleanly. A team that benchmarks Argon in October at parity with OpenAI and then ships to production in January will be paying twice the OpenAI rate for the same token counts, unless Google changes course again.
| Monthly workload (no caching) | Argon, intro rate | Argon, from Dec. 1 | Change |
|---|---|---|---|
| 100M input + 20M output | $400 | Not listed | Not listed |
| Mid-size workload, input + output | Not listed | Not listed | Not listed |
| Large workload, input + output | Not listed | Not listed | Not listed |
These figures use list prices without caching, so they overstate the pain for cache-heavy agents and understate nothing. The practical rule is to budget Argon at $4 and $20 from the start, treat the intro period as a free evaluation window, and run your cost model at both rates before committing.
Google also said, in its Google Cloud announcement of October 22, that OpenAI’s GPT-6.1 Sol, GPT Astra and GPT-6.0 Luna are in private preview for selected enterprise customers through Model Garden. That is an OpenAI listing, not a Gemini one, but it signals that cloud buyers can soon consolidate Google and OpenAI spending on one invoice.
Claude Sonnet 5.5: cheaper cache reads, a write premium and three clouds
Anthropic describes Sonnet 5.5 as a faster, lower-cost companion to Opus 5.5. Its own wording says the model runs more than 30% faster than Sonnet 5 and costs up to 30% less for most work. Opus 5.5, according to the same pages, leads in agentic coding and knowledge work and costs 40% less to run than Opus 5 on typical workloads. Haiku 5.5, released October 7, targets summarization, subagents and browser use.
Access is broad. Sonnet 5.5 is available through the Claude Platform and through Amazon Web Services, Google Cloud and Microsoft Foundry. For developers who want to build against it, our walkthrough of the Claude Agent SDK with Sonnet 5.5 covers the setup.
The pricing story has three moving parts. The base rate is reportedly $2 input and $10 output per million tokens. Cache reads fell to $0.10 on October 7. And the Anthropic pricing page offers 50% off for batch processing, with another source reporting that US-only inference is priced at 1.1 times standard rates. That last number sits on a single source, so confirm it before you route regulated workloads.
Anthropic also tied subscriptions to API spend. According to a single report of its October 7 announcement, Max 5x subscribers receive $100 per month in Claude Platform credits and Max 20x subscribers receive $200, while Team plans can pool up to $500. If that holds, a developer already paying for Max gets a small API allowance on top of the chat product, which changes the effective price of prototyping.
Where does Sonnet 5.5 lose? Mainly on the write premium and on price opacity. Third-party rate cards disagree with each other, and Anthropic’s own table was not reachable by our crawler. Nothing in our sources suggests Sonnet 5.5 costs more than its peers, but the burden of proof sits with the buyer: read the vendor page on the day you sign.
GPT-6.1 Sol: the mid-tier of OpenAI’s GPT-6 ladder
OpenAI describes GPT-6.1 Sol as an upgrade to GPT-6 Sol, aimed at agentic coding, computer use and professional work. Its API price is $2.00 per million input tokens, $0.10 per million cached input tokens and $10.00 per million output tokens, per OpenAI’s pricing page updated October 21. Standard processing can drop by half through the Batch API, which runs tasks asynchronously over 24 hours, and the same page lists a 10% uplift for data residency.
OpenAI also announced GPT-6 Sol and GPT-6 Luna on October 21, with price cuts it characterized as 50% below the GPT-5.6 promotional prices. The overlap between GPT-6 Sol and GPT-6.1 Sol is confusing, and OpenAI did not announce a retirement of the former. If your code points at one identifier today, check which one you are actually calling. For the premium end of the ladder, see our comparison of GPT-6 Astra, Gemini 4 Argon and DeepSeek.
On the consumer side, OpenAI documented three ChatGPT Pro tiers on October 22: Pro 100 at $100 per month, Pro 200 at $200 and a new Pro 500 at $500. Customers grandfathered under the previous Pro 200 allowance can keep it through October 29, after which the subscription price stays at $200 but the included usage drops. Those plans do not change API pricing, but they matter for teams that prototype in ChatGPT before moving to the API.
OpenAI’s distribution move is the bigger strategic story. On October 21 the company announced that its frontier models and Codex workflows would be available through Google Cloud’s Model Garden, with purchases transacted through Google Cloud Marketplace. For a company with committed Google Cloud spend, that turns OpenAI usage into a drawdown against an existing contract.
The honest weak point is rollout clarity. One OpenAI note says GPT-6 was launching across ChatGPT’s free and paid plans on October 19, while the October 21 Sol and Luna post says both models were in ChatGPT Work and Codex but not yet in Chat. Those statements may describe different variants or stages. Developers who depend on ChatGPT-side behavior should test against the API itself and not the chat product.
DeepSeek V4.1-Flash: the cheap outlier and its caveats
DeepSeek released V4.1-Flash on September 10, 2026, with native multimodal support and a reported one-million-token context window. On October 17 it updated the checkpoint behind the deepseek-flash alias to V4.1-Flash-1017. SiliconANGLE reported the update as a faster version of the same model, and API prices and call behavior were reported unchanged.
The price gap is the headline. If the reported peak rates of $0.30 input and $1.20 output hold, DeepSeek costs about one-seventh of the $2 input rate and roughly one-eighth of the $10 output rate that the three Western vendors charge. DeepSeek’s off-peak rate is 50% of peak, so scheduled batch-style work can drop further. Teams that can run overnight jobs get a built-in discount without a separate batch API.
There are trade-offs. The alias moves with each checkpoint, so identical prompts can return different results after an update. DeepSeek is also set to charge for retrieval: its Search and Fetch APIs cost $0.75 per 1,000 successful requests each, with billing scheduled to begin November 1, Beijing time, as Yahoo Finance reported. Calls made through DeepSeek’s own harness, Claude Code or OpenAI Codex are reportedly free until then. For budget-tier matchups, our DeepSeek, Haiku 5.5 and Gemini Flash comparison covers the cheaper siblings.
Data governance is the other question. DeepSeek is a Hangzhou-based company, and enterprises with strict residency or procurement rules will treat that as a gating factor before they look at price. We do not rank DeepSeek against the other three for regulated work, because the right answer depends on your compliance team, not on a rate card.
Where DeepSeek fits cleanly is high-volume, tolerant-of-variation work: classification, bulk summarization, extraction, first-pass triage. In those jobs, a seven-fold price gap buys room for a second verification pass with a more expensive model and still comes out ahead.
Benchmarks: what each vendor publishes, and why we skip a single ranking
A fair benchmark table needs the same tests, the same settings and the same date across all four models. We could not find that in public sources as of October 22. What exists is vendor-run evaluation, each with its own harness and its own choice of rivals.
| Source | What it reports | Why it is not directly comparable |
|---|---|---|
| OpenAI (GPT-6.1 Sol and GPT-6 release pages) | Results on DeepSWE 1.1, OSWorld 2.0 and Terminal-Bench Science 0.1, plus a claim that GPT-6 Sol beats Claude Fable 5.1 at far lower cost | OpenAI chose the tests and the comparison set. Its page says GPT-6 Astra leads the models tested at 68.1% on its science-agent evaluation |
| DeepSeek (V4.1-Flash release material) | Scores on Terminal-Bench 4.0, DeepSWE, a text-only subset of HLE, ZeroBench and AutomationBench | Footnotes say Terminal-Bench 4.0 ran with the Claude Code framework and DeepSWE with Mini-SWE-V2, and ZeroBench used Pass@5 |
| Anthropic (Sonnet 5.5, Opus 5.5 pages) | Speed and cost deltas versus earlier Claude models, plus a claim that Opus 5.5 leads in agentic coding and knowledge work | Relative to Anthropic’s own prior models, so it says little about rivals |
Each vendor reports the work it wants to be judged on. OpenAI leans on long-horizon software and computer-use tasks, DeepSeek publishes agent and multimodal evaluations with the harness named in the footnotes, and Anthropic frames gains as speed and cost against its own previous generation. That is useful context, but it does not produce a leaderboard. For a cross-vendor ranking, our LLM leaderboard covering Opus 5.5, GPT-6.1 Sol and Gemini 4 Argon is a better starting point than any single vendor page.
Why does this matter for a pricing article? Because cost per token is only half of cost per outcome. A model that needs two attempts to finish a task at $2 input costs more than a model that finishes in one at $4. The right metric is cost per accepted result on your own task set, and no public table can supply it.
Build a small evaluation harness from 50 to 100 real tasks, record pass rates and token usage per model, and divide. That number, not a leaderboard position, decides whether the November 30 Argon increase hurts or not.
Five worked cost scenarios, with the arithmetic
We modeled five monthly workloads using list prices only, so you can swap in your own volumes. The table covers the two vendors whose rates we confirmed on official pages. Outside the table, we apply the same math to Sonnet 5.5 and DeepSeek with the hedges noted earlier. For wider context on how list prices have moved, the Tokely price-trend analysis cited above is worth a second look.
| Scenario (per month) | GPT-6.1 Sol | Argon, intro rate | Argon, from Dec. 1* |
|---|---|---|---|
| 1. Support bot: no caching | Not listed | Not listed | Not listed |
| 2. Agent loop: fresh and cached input, output | Not listed | Not listed | Not listed |
| 3. Overnight job: 1,000M input, 100M output (standard) | Not listed | Not listed | Not listed |
| 3b. Same job on OpenAI’s Batch API (50% off) | $1,500 | Not stated | Not stated |
| 4. Long-context review: input and output | Not listed | Not listed | Not listed |
| 5. Coding agent: mostly cached input, heavy output | Not listed | Not listed | Not listed |
*Scenarios 2 and 5 assume the 95% cache discount carries over to the new $4 input rate, which makes cached input $0.20. Google has not restated the cached rate for the post-November price in the material we reviewed. Scenario 4 ignores any long-context surcharge, since none appeared in our sources.
Scenario 5 shows the pattern that matters most. A coding agent with 95% of its input cached still pays most of its bill on output tokens. Once caching is working, output volume sets the invoice. That is why a vendor with identical input rates can still differ on cost if its model writes longer answers or retries more often.
Apply the same volumes to Sonnet 5.5 at the reported $2 and $10 rates and scenarios 1 and 3 come out identical to GPT-6.1 Sol. Scenarios 2 and 5 diverge by the cost of cache writes, which depends on how often your prefix changes. Apply them to DeepSeek V4.1-Flash and every scenario comes out far cheaper at its reported rates, and off-peak rates are 50% of peak rates, so flexible workloads cost less when scheduled off-peak.
Those DeepSeek totals are an order of magnitude below the others. They are also the least certain numbers in this article, because they rest on a single price report. If you plan around them, pull DeepSeek’s pricing page the same day.
Five real-world moves that show how vendors are packaging these models
Pricing pages tell you the list rate. Vendor behavior over the last three weeks tells you where discounts are heading. Five concrete moves stand out.
- Anthropic halved Sonnet 5.5 cache reads on October 7. The company paired the cut with the Haiku 5.5 launch and said it lowers typical agentic-work costs by about 20%. That is a price cut aimed squarely at agent builders.
- Google set an expiry date on Argon’s launch price. The $2 and $10 rates end November 30, with $4 and $20 after. The structure rewards early adopters who lock in usage during the intro window, and penalizes those who wait.
- OpenAI put its models inside Google Cloud. The October 21 partnership announcement lets Google Cloud customers buy OpenAI models and Codex workflows through Model Garden and Google Cloud Marketplace, on consolidated billing.
- DeepSeek started charging for retrieval. Search and Fetch cost $0.75 per 1,000 successful requests each, starting November 1, after a promotional period. Model tokens are cheap, but the agent’s tools now carry their own meter.
- OpenAI cut GPT-6 prices and reshuffled Pro plans. The October 21 announcement described 50% reductions from GPT-5.6 promotional prices for GPT-6 Sol and Luna, and the three Pro tiers ($100, $200 and $500) were documented on October 22.
Taken together, these moves show vendors competing on four fronts: cached-token price, promotional expiry, distribution through cloud marketplaces, and the cost of tools around the model. A procurement team that only compares the headline input rate will miss most of the action.
There is a second lesson in the timing. Four vendors changed pricing, packaging or distribution within about three weeks of each other. Whatever you sign in late October can look stale by Thanksgiving. Favor contracts and architectures that let you swap models without rewriting your application.
Which API fits which job: six use-case recommendations
No single API wins every workload. These recommendations follow the price and packaging facts above, and they assume you validate quality on your own tasks. For readers who care about chat products more than APIs, our broader ChatGPT vs Claude vs Gemini comparison covers the consumer side.
- Long-running coding agents with stable prefixes: pick Claude Sonnet 5.5 or GPT-6.1 Sol. Both have a $0.10 cache read, and Anthropic’s agent-focused positioning suits tool-heavy loops. Compare cache write behavior against your prefix churn before choosing.
- Cloud-committed enterprises: pick the model your existing marketplace already carries. Claude runs on AWS, Google Cloud and Microsoft Foundry. OpenAI models are entering Google Cloud Model Garden. Argon sits in Google’s own stack. Drawing down an existing commitment can beat a lower list price elsewhere.
- Prototyping before Nov. 30: use Gemini 4 Argon at $2 and $10 as a low-cost evaluation window, and model your production budget at $4 and $20.
- High-volume classification, extraction and triage: test DeepSeek V4.1-Flash, and use its off-peak schedule for batchable work. Verify output stability across checkpoint updates.
- Overnight document pipelines: use a vendor with a confirmed 50% batch discount (OpenAI and Anthropic both list one) before you consider real-time pricing at all.
- Regulated or residency-sensitive data: start from your compliance requirements, not from price. OpenAI lists a 10% data-residency uplift, and Anthropic reportedly charges 1.1 times standard rates for US-only inference. Budget for those surcharges up front.
A pattern worth copying is tiered routing. Send easy requests to a cheap model, escalate hard ones to a mid-tier model like Sonnet 5.5 or GPT-6.1 Sol, and reserve the top tier for the small share of tasks that need it. With list prices this close among the three Western vendors, the routing rule saves more than the vendor choice does.
Migration guide: switching providers without breaking production
Moving an LLM workload between vendors takes more than swapping an endpoint. Prompts, tool schemas, caching and token accounting all differ. This seven-step plan keeps the risk small. If you are moving onto Claude, our Sonnet 5.5 API setup guide covers the first calls, and for DeepSeek the V4.1 Flash API setup walkthrough does the same.
- Freeze a baseline. Log 50 to 100 real production requests with inputs, outputs, token counts, latency and a pass or fail label from your own reviewers.
- Put a thin adapter in front of every vendor SDK. Your application should call one internal function, and only the adapter should know provider names, model IDs and cache parameters.
- Pin model identifiers. DeepSeek’s
deepseek-flashalias moves to the latest checkpoint, so record the checkpoint date in your logs and re-run evaluations after each update. - Port the prompt, then the cache. Keep the stable prefix first and the variable content last on every vendor, then confirm cache hits through each vendor’s usage fields before you trust the cost model.
- Shadow-run 5% to 10% of traffic. Send copies of live requests to the new model, compare pass rates and cost per accepted result, and keep the old model serving users.
- Cut over in stages with a rollback switch. Move 10%, then 50%, then 100%, with a feature flag that reverts instantly.
- Recheck the bill after two weeks. Compare actual spend against your model, and look for output-token drift, retries and tool charges.
The snippet below is a cost estimator you can adapt. It uses the confirmed GPT-6.1 Sol and Gemini 4 Argon rates, and leaves the Sonnet and DeepSeek base entries as placeholders you must fill from the vendor pages.
RATES = {
# USD per 1M tokens: fresh input, cached input, output
"gpt-6.1-sol": {"in": 2.00, "cached": 0.10, "out": 10.00},
"argon-intro": {"in": 2.00, "cached": 0.10, "out": 10.00}, # through Nov 30
"argon-from-dec-1": {"in": 4.00, "cached": 0.20, "out": 20.00}, # cached rate is an assumption
"sonnet-5.5": {"in": None, "cached": 0.10, "out": None}, # fill from Anthropic pricing page
"deepseek-flash": {"in": None, "cached": None, "out": None}, # fill from DeepSeek pricing page
}
def monthly_cost(model, fresh_in_m, cached_in_m, out_m):
r = RATES[model]
if None in (r["in"], r["cached"], r["out"]):
raise ValueError(f"Fill in rates for {model} first")
return fresh_in_m * r["in"] + cached_in_m * r["cached"] + out_m * r["out"]
# Scenario 5: coding agent, 50M fresh + 950M cached input, 100M output
print(monthly_cost("gpt-6.1-sol", 50, 950, 100)) # 1195.0
print(monthly_cost("argon-from-dec-1", 50, 950, 100)) # 2390.0Keep the adapter honest by logging estimated cost next to actual billed cost per request. A gap of more than a few percent usually means a cache miss, an unaccounted tool call, or a token-counting difference between vendors.
Pros and cons of each API
| Model | Strengths | Weaknesses |
|---|---|---|
| Claude Sonnet 5.5 | Cache read cut to $0.10 on Oct. 7. Available on the Claude Platform plus AWS, Google Cloud and Microsoft Foundry. Anthropic cites 30%+ speed gain and up to 30% lower cost versus Sonnet 5 | Cache writes carry a premium ($2.50 per 1M). Base rate table was not directly verifiable by our crawler, and third-party pages disagree |
| GPT-6.1 Sol | Fully confirmed rate card: $2 / $0.10 / $10. 50% Batch API discount. Sits between Astra and Luna so you can route up or down within one vendor | Naming overlap with GPT-6 Sol. Mixed signals on ChatGPT rollout. Data residency adds 10% |
| Gemini 4 Argon | Intro price matches OpenAI at $2 / $10 with a 95% cache discount. Wide Google surface: API, AI Studio, Antigravity, AI Mode | Price doubles to $4 / $20 after Nov. 30. Cached rate after the change is not restated in our sources |
| DeepSeek V4.1-Flash | Reported peak rates far below the others, with off-peak at 50% of peak. Native multimodal support and a reported 1M-token context | Moving alias changes behavior between checkpoints. Search and Fetch now billed from Nov. 1. Provenance and data-residency questions for regulated work. Price figures rest on limited sources |
The table hides one asymmetry. Three of the four vendors compete mainly on stability and ecosystem, since their list prices are so close. The fourth competes on cost. Those are different buying decisions, and mixing them into one ranking produces a misleading number.
Verdict: where Gemini, ChatGPT and Claude land on API cost
On the data we can confirm, the three Western APIs tie on cached-input price today, so the purchasing decision rests on three things: how well your prefix caches, how much output your workload generates, and what happens to Argon’s price on December 1. If you want a wider view before deciding, the AI and machine learning section tracks each vendor’s updates.
- Best for predictable budgets: GPT-6.1 Sol. It has the only fully confirmed rate card among the four, a 50% batch option and a clear tier ladder above and below it.
- Best for agent builders who control cache hygiene: Claude Sonnet 5.5. The October 7 cache cut and cloud breadth are strong points, with the caveat that you should read Anthropic’s live pricing page and plan for the write premium.
- Best for short-term evaluation: Gemini 4 Argon until November 30. After that date, it costs twice as much as GPT-6.1 Sol on list price, and it needs to earn that premium on quality.
- Best for volume and cost: DeepSeek V4.1-Flash, with governance review first. If the reported rates hold, it costs roughly one-seventh of the others on input and one-eighth on output.
Our call for a typical engineering team: default to GPT-6.1 Sol or Sonnet 5.5, whichever passes your evaluation set at the lower cost per accepted result. Budget Argon at $4 and $20, treat the October price as a trial, and add DeepSeek as a cost tier for work that tolerates variation. Revisit the decision on December 1, when the first real Argon invoices land.
FAQ
What does GPT-6.1 Sol cost per million tokens?
OpenAI’s pricing page, updated October 21, 2026, lists GPT-6.1 Sol at $2.00 per million input tokens, $0.10 per million cached input tokens and $10.00 per million output tokens. The Batch API cuts standard processing by 50%, and data residency adds 10%.
When does Gemini 4 Argon’s price go up?
Google’s introductory rates of $2 per million input tokens and $10 per million output tokens run through November 30, 2026. From December 1, the public price becomes $4 input and $20 output per million tokens, which doubles both rates.
How much is a cached token on each API?
GPT-6.1 Sol lists cached input at $0.10 per million tokens. Gemini 4 Argon prices cached input 95% below its input rate, which is $0.10 at the introductory $2 rate. Anthropic cut Claude Sonnet 5.5 cache reads from $0.20 to $0.10 per million tokens on October 7, 2026, and lists cache writes at $2.50.
Is Claude Sonnet 5.5 the newest Claude model?
Sonnet 5.5 launched on September 28, 2026 as part of the Claude 5.5 family. Claude Haiku 5.5 followed on October 7, and Anthropic also lists Claude Opus 5.5 as a current model that leads in agentic coding and knowledge work.
Is DeepSeek V4.1-Flash cheaper than the others?
Reported peak rates put it far below the three Western vendors, and DeepSeek’s off-peak rates are 50% of peak. The exact numbers vary between pages, so check DeepSeek’s pricing documentation before budgeting. Its Search and Fetch tools also cost $0.75 per 1,000 successful requests each from November 1.
Can I buy OpenAI models through Google Cloud?
OpenAI announced on October 21, 2026 a partnership to offer its frontier models and Codex workflows through Google Cloud’s Model Garden, with purchases through Google Cloud Marketplace. Google Cloud described GPT 6.1 Sol, GPT Astra and GPT 6.0 Luna as available in private preview to selected enterprise customers.
Which API is best for AI agents?
It depends on cache hit rate and output volume. Sonnet 5.5 and GPT-6.1 Sol share the same $0.10 cache-read rate, so evaluate both on your own task set and compare cost per accepted result. Budget Gemini 4 Argon at its post-November rates.
