One road stays level the whole way. The other steps up once you pass a certain distance. That is the difference the price list does not show you.
Almost every Claude-versus-Gemini comparison answers the question "which model is smarter". It is the wrong question for a business, partly because the answer changes every few weeks, and mostly because for the work most companies actually do, both are good enough that the benchmark never decides anything.
The question worth answering is narrower and duller: which one is cheaper for the shape of work you are actually running? That answer is stable, it is knowable from published pricing, and it is not what the headline rate tells you.
A disclosure before anything else: Anchor Sprint is a member of the Anthropic Claude Partner Network, so we have a stated preference. We have tried to be straight about where Gemini wins — there are several places, and one of them is significant. Weigh our position accordingly.
Where Gemini genuinely wins
Getting this out of the way first, because a comparison that finds no merit in the other side is marketing, not analysis.
There is a real free tier. Gemini offers free-of-charge usage on the Flash models and on Gemini 2.5 Pro. Anthropic does not offer a free tier on the Claude API. If you are prototyping, teaching a team, or running something genuinely low-stakes, that difference is not marginal — it is the difference between a bill and no bill.
Search grounding is built in and cheap. Gemini includes grounding with Google Search at 5,000 free search requests per month, shared across the Gemini 3.x models, then $14 per 1,000 requests. If your task is fundamentally "answer using current information from the web", that is a first-party capability wired into the model's own pricing, and it is Google's home turf.
The cheap end is genuinely cheap. Gemini 3.5 Flash-Lite runs $0.30 per million input tokens and $2.50 output. Gemini 3.6 Flash is $1.50 and $7.50. For high-volume, low-judgement work these are aggressive numbers.
And the one nobody at a vendor likes to admit: you may already be paying for it. Google Workspace penetration among Malaysian and Singaporean businesses is high. If your company already lives in Gmail, Drive and Docs, Gemini reaches your data with no integration work at all. That is not a model-quality argument. It is often the deciding one anyway, and pretending otherwise would be dishonest.
The 200k cliff
Now the structural difference, which is where this gets interesting.
Gemini prices in two context tiers. For Gemini 3.1 Pro, published rates are:
- Prompts up to 200k tokens: $2.00 per million input, $12.00 output
- Prompts over 200k tokens: $4.00 per million input, $18.00 output
Input doubles. Output rises by half. There is a step in the road.
Claude prices one flat rate across its entire window. Claude Opus 5 is $5 input and $25 output per million tokens whether the prompt is two thousand tokens or nine hundred thousand. Claude Sonnet 5 is $2 and $10 on current introductory pricing — flat. Fable 5, Opus 5 and Sonnet 5 all carry a 1 million token context window at a single rate.
So the comparison inverts depending on how much context you send:
- Short prompts. Gemini 3.1 Pro at $2 input is level with Sonnet 5's introductory $2 and well under Opus 5's $5. Gemini wins on price.
- Long prompts. Gemini 3.1 Pro at $4 input is double Sonnet 5 and close to Opus 5 — while Claude has not moved at all.
The headline "$2 vs $5" that most comparison tables print is true only in the first case.
One honest note: OpenAI does the same thing — GPT-5.6-sol runs $5 input on short context and $10 on long, which we work through in Claude vs ChatGPT. Context-tiered pricing is the industry norm; Anthropic's flat rate across 1M is the outlier. This is a genuine structural difference, not a Gemini failing.
Why this hits agents specifically
For a chat assistant, the 200k cliff is mostly theoretical. People do not paste 200,000 tokens into a chat box.
Agents cross it constantly, and not because anyone designed them to. An agent accumulates context as it works — the system prompt, the tool definitions, the results of every tool call it has already made, the running history of its own reasoning. A support-triage agent that starts at 8k tokens on turn one can be well past 200k by turn thirty of a complicated case. Nobody chose that. It is just what carrying state looks like.
Which means the pricing tier you are actually billed at is decided by the length of the session, not the model you picked. A workload that looks cheaper on the price list can quietly bill at the higher tier for most of its real work, and you will not see it in a benchmark or a demo. You will see it on the invoice.
This is the single most useful thing we can tell you in this article: when comparing for agent work, price the long-context tier, not the short one. Then compare against Claude's flat rate. The gap is usually much smaller than the marketing suggests, and sometimes it points the other way.
You can run your own numbers against both structures in our Claude cost calculator.
Two different philosophies of caching
Both vendors let you cache a stable prompt prefix so you are not paying full rate for identical bytes thousands of times a day. They charge for it differently, and the difference suits different workloads.
Claude charges cache hits at 10% of base input — an Opus 5 cache hit is $0.50 against $5 base. Cache writes cost 1.25× base for a five-minute window and 2× for a one-hour window. You pay more up front to write, then very little to read, and nothing for the time in between.
Gemini charges context caching at $0.20 per million for prompts under 200k and $0.40 over — a similar ratio to base — plus a storage price of $4.50 per million tokens per hour.
That hourly storage charge is the thing to notice. If you cache a large document and hit it constantly for ten minutes, Gemini's model is fine. If you cache it and hit it sporadically across a working day, you are paying rent on it the whole time. Claude's higher write cost and zero storage charge suits long-lived, intermittently-used caches; Gemini's suits short, intense bursts.
Neither is better. They are priced for different shapes, and knowing which shape you have is worth more than knowing which vendor is cheaper.
One more billing difference worth knowing
Gemini's published output price is explicitly labelled "including thinking tokens". The reasoning the model does before answering is billed as output.
This is not a criticism — it is transparent and it is in the pricing table. But it does mean that on reasoning-heavy tasks, Gemini's effective output cost is higher than the headline number implies, in a way that varies with how hard the model decides to think. Budget with headroom on anything analytical.
How we actually choose
Across the builds in our agent use-case directory, the decision rarely comes down to model quality. It comes down to four things:
Where does the data already live? If it is in Google Workspace and it needs to stay there, Gemini's integration advantage is real and you should weigh it heavily. Fighting your own stack to use a marginally preferred model is a bad trade.
How long are the sessions? Short, stateless calls favour Gemini's cheaper short-context tier. Long agent runs that accumulate context favour Claude's flat rate. This is the axis most comparisons miss entirely.
Does the task need live web information? Gemini's built-in Search grounding is a genuine advantage and it is priced reasonably.
What happens when it is wrong? This is the one we care about most, and it is model-agnostic. If output flows into another system unreviewed, buy the more capable tier — of either vendor. We wrote about that trade-off within the Claude range in Claude Opus 5 vs Sonnet 5, and the logic transfers.
A short decision list
- Already deep in Google Workspace, with data that stays there? Gemini, and do not overthink it.
- Short, stateless, high-volume calls? Gemini Flash or Flash-Lite. The cheap end is genuinely cheap.
- Long-running agents that accumulate context? Price Gemini's over-200k tier honestly, then compare to Claude's flat rate. Claude usually wins this one, and it is the reason we build on it.
- Task depends on current web information? Gemini's Search grounding.
- Prototyping with no budget? Gemini's free tier. There is no Claude API equivalent.
- Output flowing unreviewed into a ledger, a CRM or a customer's inbox? Buy capability, from either vendor. Do not optimise this one on price.
The honest summary
For short prompts Gemini is cheaper. For long-context agent work the gap narrows sharply or reverses, because Gemini's rate steps up past 200k tokens and Claude's does not. Gemini has a real free tier, native Search grounding and an integration advantage in any Google Workspace shop — all three are genuine and none of them are about which model is smarter.
Pick on the shape of your workload and where your data lives. The benchmark scores will have changed by the time you finish the migration anyway.
Free consultation
Not sure which side of the 200k line your workload sits on?
Tell us what the agent does and roughly how long a session runs. We will price it both ways and tell you plainly which vendor is cheaper for your shape of work — including when the answer is not the one we sell.
On fairness and timing. All pricing here was taken from Anthropic's and Google's own published documentation on 11 August 2026. AI pricing changes frequently — verify current rates with each vendor before committing budget. Anchor Sprint is a member of the Anthropic Claude Partner Network and has a disclosed preference for Claude. We are not affiliated with, authorised by, or endorsed by Google. Gemini and Google Workspace are trademarks of Google LLC; Claude and Anthropic are trademarks of Anthropic, PBC.
Compare all three: ChatGPT vs Claude vs Gemini for Malaysian business