From 1 December 2026 these two models cost exactly the same: $4 per million input tokens, $20 per million output. Google’s introductory rate halves that until 30 November, and that is the only price advantage either side has.
Pick Claude Opus 5.5 for production work you are committing to: it leads the one independent benchmark by five points, its caching discount is contractual rather than opportunistic, and it has a published retirement date.
Pick Gemini 4 Argon for the next six weeks of evaluation, for workloads that need a single response longer than 128,000 tokens, or if your stack already runs on Google Cloud.
The 30-second answer
Pick Gemini 4 Argon if: you want frontier capability at mid-tier cost before 1 December; you need single responses above 128K tokens; you need Google Search or Maps grounding in-model; your data already lives in Google Cloud.
Pick Claude Opus 5.5 if: you are deploying rather than evaluating; you want the highest independently measured capability at this price; your workload re-sends a large stable prefix and you need the cached rate to be a guarantee; procurement needs a retirement date in writing.
Both run a 1M-token context window at standard rates with no long-context surcharge, which is the one thing that is genuinely a tie. See long-prompt API costs for where the rest of the field falls off a cliff.
The rate cards, side by side
Prices verified 15 October 2026 against Google’s Gemini API pricing page and Anthropic’s pricing page.
| Per million tokens | Gemini 4 Argon | Claude Opus 5.5 |
|---|---|---|
| Input — now | $2.00 | $4.00 |
| Input — from 1 Dec 2026 | $4.00 | $4.00 |
| Output — now | $10.00 | $20.00 |
| Output — from 1 Dec 2026 | $20.00 | $20.00 |
| Cached input read | $0.10 → $0.20 | $0.20 |
| Cache write | n/a — no write fee | $5.00 (5 min) / $8.00 (1 hr) |
| Caching mechanism | implicit only | explicit, pinned |
| Batch — from 1 Dec | $2.00 / $10.00 | $2.00 / $10.00 |
| Context window | 1,048,576 | 1,000,000 |
| Long-context surcharge | none | none |
| Max output | 65,536 default; 1,048,576 long decode | 128,000; 300,000 on batch (beta) |
| Faster tier | Priority: $7.20 / $36.00 from 1 Dec | Fast mode: $8.00 / $40.00 |
| Published retirement date | none | not sooner than 22 Sep 2027 |
| Minimum notice before retirement | none published | 60 days |
Two exact collisions in that table are worth naming. The standard rate from December is identical. The cached read price is identical at $0.20. Everything that separates these models on cost sits in the rows around them — and three of those rows are worth real money.
The gap the price hides
When Argon went on sale on 14 October, the capability case for it was entirely first-party: 77.9% on DeepSWE v1.1, 51.3% on AutomationBench, 91.7% on LVBench, 68% on CWE-bench v1. Good numbers, published by the company selling the model, benchmarked on the Fairwind build rather than the one customers now call.
Independent scores have since arrived, and they are less flattering:
| Model | Artificial Analysis Intelligence Index | Setting tested |
|---|---|---|
| Claude Opus 5.5 | 58 | max with fallback |
| Claude Sonnet 5.5 | 56 | max with fallback |
| Claude Opus 5.5 | 56 | xhigh |
| Claude Fable 5.1 | 53 | max with fallback |
| GPT-6 Astra | 53 | max |
| Gemini 4 Argon | 53 | high |
| GPT-6.1 Sol | 52 | max |
| Grok 4.7 | 46 | xhigh |
The setting column matters and it does not rescue Argon. Google’s docs list Argon’s thinking levels as low, medium and high, so high is Argon’s ceiling — there is no setting above it to try. Opus 5.5 reaches 58 at its top setting and still scores 56 at xhigh, three points clear of Argon’s maximum.
Read plainly: Argon lands in a three-way tie with GPT-6 Astra and Claude Fable 5.1, two points behind Anthropic’s mid-tier Sonnet 5.5, and five behind the model it will be price-matched against. That is a real improvement on Google’s seventh-place standing among labs when it had nothing at the frontier to rank. It is not a lead.
The output ceiling is a cost difference at identical prices
The specification that moves money here is not the context window — both are 1M, both at standard rates. It is the output ceiling, because output is the expensive half of both rate cards.
| Max single response | Cost of one maximum-length response (Dec rates) | |
|---|---|---|
| Argon, default | 65,536 | $1.31 |
| Opus 5.5, synchronous | 128,000 | $2.56 |
Opus 5.5, batch with output-300k beta | 300,000 | $3.00 |
| Argon, long decode | 1,048,576 | $20.97 |
Argon’s headline capability — “long decodes up to 1M tokens” — is also the only way either of these models can bill twenty dollars for one API call. Enabling it raises your worst-case single call 16x against Argon’s own default and 8.2x against Opus 5.5’s synchronous ceiling, at an identical per-token price.
The practical rule is the same on both models and more urgent on one: set an explicit maximum-output value in code before an agent loop goes to production. On Opus 5.5 the default ceiling caps the damage at $2.56. On Argon with long decode enabled, there is effectively no cap worth relying on.
Caching: a guarantee against a lottery
Both vendors charge $0.20 per million for a cached input read from December. They sell completely different things for that price.
Anthropic sells a guarantee. You mark a prefix, pay a write fee — $5 per million for the 5-minute cache, $8 for the hour — and every subsequent read of that prefix costs $0.20 per million. Minimum cacheable prompt is 512 tokens. The write fee is the price of certainty: it costs 1.25x the base input rate, so caching pays for itself from the first read.
Google sells a probability. Argon supports implicit caching only. Google applies the discount when it recognises a repeated prefix; you cannot pin one, reserve one, or guarantee a hit rate. There is no write fee, so there is no downside to a miss — you simply pay the full $4 input rate. There is also no floor.
That makes the comparison a straight expected-value question. Per million tokens of repeated prefix, from 1 December:
| Implicit hit rate on Argon | Argon cost /M prefix | Opus 5.5 explicit cache, steady state |
|---|---|---|
| 100% | $0.20 | $0.20 |
| 90% | $0.58 | $0.20 |
| 75% | $1.15 | $0.20 |
| 50% | $2.10 | $0.20 |
| 25% | $3.05 | $0.20 |
| 0% | $4.00 | $0.20 |
Anthropic’s column does not move, which is the entire product. Google’s only matches it at a hit rate of 100%, and that is a number no one — including Google — publishes for your workload.
What a month actually costs
Same rate card, different bill. A concrete agentic workload, priced at December rates: 10,000 calls a month, each sending 50,000 input tokens of which 40,000 are a stable prefix, and returning 15,000 output tokens. That is 500M input tokens (400M of it prefix) and 150M output.
| Configuration | Monthly bill |
|---|---|
| Either model, no caching | $5,000.00 |
| Opus 5.5, 1-hour cache rewritten hourly | $3,707.76 |
| Opus 5.5, cache kept warm by traffic | ~$3,480.00 |
| Argon, implicit caching at a 90% hit rate | $3,632.00 |
| Argon, implicit caching at a 50% hit rate | $4,240.00 |
| Argon, promotional rates, 50% hit rate (until 30 Nov) | $2,120.00 |
The break-even implicit hit rate against the hourly-rewrite case is about 85%. Against the warm-cache case — which is what 13.7 calls an hour actually produces, since reads refresh an Anthropic cache’s lifetime — Argon needs essentially perfect implicit caching to draw level.
Until 30 November, none of that matters: Argon is roughly 43% cheaper than a well-cached Opus 5.5 on this workload, because the promotional rate halves both input and output. From 1 December, at list prices that are character-for-character identical, the better-caching model wins by 15-20% on a bill like this one. Price your own shape in the API cost calculator — the tokenizer difference between these two vendors alone moves cross-vendor comparisons by about 30%.
Lifecycle: a date against no date
For anything you are committing a codebase to, this is the section that should decide it.
Claude Opus 5.5 is published as retiring “not sooner than 22 September 2027”, and Anthropic commits to at least 60 days’ notice before any publicly released model retires. Those dates apply to the Claude API, Claude Platform on AWS and Microsoft Foundry; Amazon Bedrock and Google Cloud set their own. Anthropic has honoured the floor narrowly before — Sonnet 4.5 got 61 days — but it honoured it.
Gemini 4 Argon does not appear on Google’s deprecation page at all. Google publishes no minimum notice period; its policy states that listed shutdown dates “indicate the earliest possible dates on which a model might be retired” and that it will “communicate the exact shutdown date to users with advance notice” — a commitment with no number attached. Five days before Argon went on sale, Google deprecated Gemini 3.5 and 3.7 Flash and rerouted the model IDs silently.
Both vendors’ current dates are tracked on the model retirement calendar.
Paying for speed
Both vendors now sell latency separately, and the prices are close enough to compare directly:
- Argon Priority tier: $3.60 / $18.00 per million now, $7.20 / $36.00 from 1 December — 1.8x the standard rate.
- Opus 5.5 Fast mode: $8.00 / $40.00 per million — 2x the standard rate, in research preview, and available on the first-party Claude API only, not on AWS, Google Cloud or Foundry.
Argon’s faster tier is marginally cheaper and more broadly available; Anthropic’s is a preview with a platform restriction. Neither is priced for volume. Both are published on their vendor’s main rate card, which is more than can be said for most premium-latency products.
Which one to pick by use case
| Workload | Pick | Why |
|---|---|---|
| Long-running agentic coding | Claude Opus 5.5 | Five points of independent capability at the same price; Claude Code is the stronger harness |
| Evaluating a frontier model this quarter | Gemini 4 Argon | Half price until 30 November; run the bake-off, price production at $4/$20 |
| Single responses over 128K tokens | Gemini 4 Argon | Long decode to 1M; nothing in Anthropic’s lineup goes past 300K, and only on batch |
| Large repeated prefix, cost-sensitive | Claude Opus 5.5 | $0.20 cached reads are guaranteed, not probabilistic |
| Research with live web or map data | Gemini 4 Argon | Search and Maps grounding in-model, 5,000 free search requests a month then $14 per 1,000 |
| Anything procurement must sign off | Claude Opus 5.5 | A published retirement date and a 60-day notice floor |
| High volume at acceptable quality | Neither | Claude Sonnet 5.5 scores 56 at $2/$10 — two points above Argon at half the December price |
That last row is the uncomfortable one. On the independent index, Anthropic’s mid-tier model outscores Google’s frontier model, and from 1 December it does so at half the price. The strongest argument against buying Argon at $4/$20 is not Opus 5.5. It is Sonnet 5.5.
The verdict
Google has built a frontier model a customer can finally buy, and priced it at exactly what Anthropic charges for a better one. The introductory window is real and worth using — six weeks of frontier-class capability at mid-tier cost is the cheapest evaluation opportunity of the quarter, and Gemini’s wider subscription story improves because of it.
But the decision that matters is the one dated 1 December. At that point Argon offers five fewer points of independently measured capability, a caching discount it cannot promise, a worst-case single call eight times larger, and no retirement date, for the same money as Claude Opus 5.5. The things Argon wins — the output ceiling, in-model grounding, Google Cloud adjacency — are real and specific. They are reasons to choose it for a workload, not reasons to standardise on it.
Evaluate Argon now. Budget for $4/$20. Then make the call on the evidence, which currently points at Anthropic.
What changed
- 15 Oct 2026: Page created. Artificial Analysis published independent scores — Argon 53 at
highagainst Opus 5.5’s 58 — and Google’s docs confirmed Argon’s output default is 65,536 with long decodes to 1,048,576. - 14 Oct 2026: Gemini 4 Argon reached general availability at $2/$10, reverting to $4/$20 — Opus 5.5’s exact rate — on 1 December 2026.
- 22 Sep 2026: Claude Opus 5.5 launched at $4/$20 against Opus 5’s $5/$25, with cache reads cut 60% to $0.20 — a default, not a discount.
Gemini 4 Argon vs Claude Opus 5.5 — frequently asked questions
Which is cheaper, Gemini 4 Argon or Claude Opus 5.5?
Argon is half the price until 30 November 2026 and exactly the same price afterwards. Google's introductory rate is $2 per million input tokens and $10 per million output; Google's own pricing page states that becomes $4 and $20 on 1 December 2026, which is Claude Opus 5.5's standing rate to the decimal. Batch is 50% off on both sides, so that collision holds there too ($2/$10 each). From December, the cheaper model is whichever one your caching and output settings make cheaper — not whichever one has the lower list price, because neither does.
Is Gemini 4 Argon as good as Claude Opus 5.5?
Not on the only independent measurement available. Artificial Analysis scores Gemini 4 Argon at 53 on its Intelligence Index, tested at `high` — which is Argon's top thinking level — against 58 for Claude Opus 5.5 and 56 for Claude Sonnet 5.5. So the independent read puts Argon five points behind Anthropic's flagship and two points behind Anthropic's mid-tier model, at the same price as the flagship from 1 December. Google's own numbers are stronger, including 77.9% on DeepSWE v1.1, but they are first-party and were published before general availability.
Why can a single Argon call cost so much more than a single Opus 5.5 call?
Because the output ceilings differ by a factor of eight. Claude Opus 5.5 caps a synchronous response at 128,000 tokens, which bills $2.56 at $20 per million. Argon defaults to 65,536 output tokens but supports long decodes up to 1,048,576 — and a maximum-length long decode bills $20.97 from 1 December. Same per-token price, 8.2x the worst-case single call. Set an explicit output cap before an agent loop reaches production on either model, and treat enabling Argon's long decode as a budgeting decision rather than a configuration one.
Is Argon's 95% cached-input discount as good as Anthropic's prompt caching?
Only if it lands nearly every time. Argon supports implicit caching only: Google applies the discount when it recognises a repeated prefix, and you cannot pin, reserve or guarantee a hit. Anthropic sells explicit caching — you pay a write fee ($5 per million for 5 minutes, $8 for an hour) and then every read costs $0.20 per million, guaranteed. The cached read prices are identical at $0.20. On the worked example below, Argon's implicit caching needs roughly an 85% hit rate just to match Anthropic's bill at identical list prices, and more than that if continuous traffic keeps the Anthropic cache warm.
Which has the safer lifecycle terms?
Claude Opus 5.5, clearly. Anthropic publishes a retirement date of 'not sooner than 22 September 2027' and commits to at least 60 days' notice before any released model retires. Gemini 4 Argon does not appear on Google's deprecation page at all, so it carries no date of any kind, and Google publishes no minimum notice period — its policy says listed shutdown dates are 'the earliest possible dates' and that it will communicate the exact date 'with advance notice', unquantified. Google also deprecated two Flash models in October 2026 and rerouted them silently, which is the precedent to price in.
Should I evaluate Argon during the promotional window?
Yes, if you have an evaluation to run — but price the production decision at $4/$20. Until 30 November a frontier-class Google model costs what mid-tier GPT-6 Sol and Claude Sonnet 5.5 cost, which makes the next six weeks unusually cheap for a real bake-off. A pilot whose business case only closes at $2/$10 is a pilot that fails on 1 December.