Gemini 3.8 Flash vs Claude Haiku 4.5: two different answers to "make it cheap and fast"

Gemini 3.8 Flash and Claude Haiku 4.5 sit at the same corner of their vendors' catalogs, and that is most of what they share. Gemini 3.8 Flash is Google's new reasoning-heavy workhorse, launched September 2, 2026, at an introductory $0.75 per million input tokens and $3.75 per million output. Claude Haiku 4.5 is Anthropic's sprint specialist, on the market since October 15, 2025, at $1 and $5 per million. Both are called "fast," but they answer a different question: Gemini 3.8 Flash gets speed by thinking hard and generating at roughly 350 output tokens per second, while Claude Haiku 4.5 gets speed by not thinking much at all and answering in under a second.

This is not the usual "new model beats old model" matchup, because the cheaper of the two on paper is the smarter one, and the pricier one has an eleven-month head start in production. If you are choosing between them today, the deciding factors are latency budget, token appetite, and how much reasoning depth your workload actually rewards.

Why Google's "flash" model no longer means what Haiku means by "fast"

The two models embody opposite design philosophies for the low-cost tier. Anthropic built Claude Haiku 4.5 to stay lean: a 200,000-token context, a 64,000-token output limit, text and image input, and a default that does not burn tokens on extended reasoning. Anthropic positioned it at launch as matching Claude Sonnet 4's coding and agent performance at about a third of the cost, and its calling card in practice is time-to-first-token: independent measurements put Claude Haiku 4.5's first token at under a second, which is what makes it work for chat, classification, extraction, and moderation pipelines where a user or a queue is waiting.

Google, by contrast, has been turning its Flash line into a thinker. Gemini 3.8 Flash supports a reasoning-effort control — low, medium, or high — and it reasons by default, with thinking tokens billed as output. On Artificial Analysis it scores 59 on the Intelligence Index in its high-reasoning mode (57 at medium, 52 at low), a level that sits near models several times its price. Claude Haiku 4.5 measures 24 out of the box on the same index, and about 37 when its optional extended-thinking mode is switched on. That gap is real, but it is partly a setting, not just a capability: comparing Gemini 3.8 Flash at high effort to Claude Haiku 4.5 at idle compares one model doing its hardest work against another deliberately doing its easiest.

Specs, side by side

• Context — Gemini 3.8 Flash: 1,048,576 tokens in, 65,536 out. Claude Haiku 4.5: 200,000 tokens in, 64,000 out.

• Price — Gemini 3.8 Flash: $0.75 / $3.75 per million (promotional through December 31, 2026, then $1.50 / $7.50). Claude Haiku 4.5: $1.00 / $5.00 per million, with cached input at $0.10.

• Output speed — Gemini 3.8 Flash: ~350 tokens per second (Artificial Analysis). Claude Haiku 4.5: ~95–105 tokens per second out of the box.

• First token — Gemini 3.8 Flash: mid-range p50 around 4.8 s on OrcaRouter telemetry. Claude Haiku 4.5: under 1 s on independent measurements.

• Reasoning — Gemini 3.8 Flash: effort levels low / medium (default) / high. Claude Haiku 4.5: off by default, optional extended thinking.

• Inputs — Gemini 3.8 Flash: text, image, video, audio, file. Claude Haiku 4.5: text and image.

• Released — Gemini 3.8 Flash: September 2, 2026. Claude Haiku 4.5: October 15, 2025.

What the benchmarks actually say, with the labels attached

Google's headline for Gemini 3.8 Flash is vendor-reported: on its own DeepSWE v1.1 software-engineering evaluation it claims 71.0–73.7%, up from 65.3% for Gemini 3.7 Flash and within reach of Claude Opus 5's 74.0%. Those are Google's numbers on Google's harness, and the honest way to read them is as an upper bound on what the model can do, not a guaranteed result in your codebase. The independent signal points the same direction: Artificial Analysis's 59 Intelligence Index is the number that has moved the market's attention, and the agency of that gain is real — Artificial Analysis attributes most of it to better tool use and agentic behavior rather than raw knowledge.

Claude Haiku 4.5's comparable claim is Anthropic's launch statement that it matches Claude Sonnet 4 on coding and agent work at a fraction of the price, plus its reported 73% on SWE-bench Verified. Independent aggregators put Claude Haiku 4.5's SWE-bench numbers considerably lower, which is normal for a small model — the vendor number reflects its best harness configuration. If you need a coding agent rather than a chat model, Gemini 3.8 Flash's independent agentic gains and 1M-token context make it the stronger choice on paper; if you need a model that reliably returns cheap, fast, well-formed answers, Claude Haiku 4.5's track record is the one that has been load-tested in production for close to a year.

The cost picture has an expiry date

Right now, the arithmetic favors Gemini 3.8 Flash: its promotional rate undercuts Claude Haiku 4.5 on both input and output tokens. That is the entire point of the promo — Google priced a smarter model below Anthropic's budget model to pull in adoption. But the promo is temporary. Google has stated that the $0.75 / $3.75 rate for Gemini 3.8 Flash lasts through December 31, 2026, and that the standard rate from January 1, 2027 is $1.50 / $7.50. Claude Haiku 4.5 has no such cliff: at $1 / $5 it will be the cheaper model on both axes from January onward, and its cache-read price of $0.10 per million is far below Gemini's.

There is a second cost trap hiding in the per-token prices, and it is worse for Gemini 3.8 Flash. Reasoning models spend output tokens thinking, and Artificial Analysis found Gemini 3.8 Flash uses roughly 30% more output tokens per task than Gemini 3.7 Flash — averaging around 48,000 output tokens per task — which pushes its real per-task cost about 40% higher than its predecessor's, even at the same per-token price. Claude Haiku 4.5 is frugal by construction: with extended thinking off, what it returns is mostly the answer. For a workload where cost per task matters more than cost per token, that difference can outweigh the sticker price.

Which one for which job

• Sub-second interactive chat, moderation, extraction, classification at high volume — Claude Haiku 4.5. Its latency profile is the point, and its 200K context covers most of these workloads.

• Long-horizon coding agents, multi-file edits, tasks that need to hold a whole repository in context — Gemini 3.8 Flash. The 1M-token context and the agentic benchmark gains are precisely what this needs.

• Video, audio, and long-file understanding — Gemini 3.8 Flash. Claude Haiku 4.5 does not take video or audio input at all.

• A workload you will re-budget in January — do the math on per-task tokens, not per-token price. Gemini 3.8 Flash at $0.75 / $3.75 is a great deal today and a different proposition at $1.50 / $7.50 with a 30% token-tax on reasoning.

Trying both from one key

Both models are hosted on OrcaRouter at the provider list price with zero markup, which makes the test-before-you-commit version of this comparison trivial. Gemini 3.8 Flash is on OrcaRouter at Google's list price; Claude Haiku 4.5 is on OrcaRouter at Anthropic's list price. Point the same OpenAI-compatible endpoint and the same key at model ids google/gemini-3.8-flash and anthropic/claude-haiku-4.5, run the identical prompt through each, and compare the actual latency, cost, and output on your own traffic instead of trusting a benchmark table. Because vendor price changes are passed through the same day, the promotional Gemini rate is live here now, and it will update again automatically when the standard rate takes effect in January — your code does not need to change, only your expectation of the bill.

The verdict depends on which "fast" you are buying. If your users are waiting on the first token and the task is bounded, Claude Haiku 4.5 remains the rational default — it is proven, lean, and sub-second. If the task is open-ended enough to reward thinking, Gemini 3.8 Flash is offering near-frontier reasoning at a promotional price that undercuts the incumbent, and the window to try it at that price closes on December 31. For most teams the cheapest correct answer is not one of them but a routing rule: send the simple, latency-sensitive calls to Claude Haiku 4.5 and the hard, context-heavy ones to Gemini 3.8 Flash — one key, one bill, and each model doing the work it is actually good at.