Claude Haiku 5.5 vs Gemini 3.8 Flash: Price, Speed, Benchmarks

Claude Haiku 5.5 and Gemini 3.8 Flash are the cheap, fast workhorses of their respective model families, and on paper they look like the same product: both advertise roughly a 1-million-token context window, both let developers dial an "effort" level up or down, and both are pitched at the same job — high-volume classification, chat, extraction, and lightweight agentic work where frontier-model pricing doesn't make sense. The gap shows up in the details. Anthropic's published pricing puts Claude Haiku 5.5 at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, while Google's Gemini API pricing page lists Gemini 3.8 Flash at a promotional $0.75 input / $3.75 output per million tokens — a rate that is scheduled to double on January 1, 2027. Claude also allows a larger 128K-token output budget against Gemini's 64K. Google's model, in turn, natively reads audio, video, and PDFs, inputs Claude's current line-up doesn't advertise.

Price per million tokens: the real gap

Price is the headline reason either model exists. Anthropic's own pricing page lists Claude Haiku 5.5 at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, rising to $0.50 input / $2.50 output above that threshold. Anthropic's launch announcement frames this as roughly a 75% average price cut from Claude Haiku 4.5 ($1 / $5 per million tokens), broken out as a 90% reduction for prompts under 100K tokens and a 50% reduction above it.

Google's Gemini API pricing page lists Gemini 3.8 Flash's standard tier at $0.75 per million input tokens and $3.75 per million output tokens — itself described as a discount from a $1.50 / $7.50 base rate that takes effect January 1, 2027. Google also publishes separate Batch, Flex, and Priority tiers, with Priority running as high as $1.35 / $6.75 (promo) and the regular Standard and Priority tiers offered on a free tier that Batch and Flex don't get.

Two things matter beyond the headline numbers. First, Claude Haiku 5.5's price is tiered by prompt length: once a single prompt crosses 100,000 tokens, every rate — input, output, and both cache rates — jumps five-fold. Anthropic's own documentation flags this as unique to Haiku 5.5 among current Claude models. Second, Gemini 3.8 Flash's $0.75 / $3.75 rate is explicitly introductory; Google's pricing page states it reverts to $1.50 / $7.50 on January 1, 2027, which roughly doubles the per-token cost without any model change. Readers budgeting a high-volume pipeline into next year should price it at the post-promo rate, not the current one.

Context window and output limits

Both companies market a 1-million-token context window, but the ceiling on what a model can hand back in one response differs meaningfully. Anthropic's models overview documentation lists Claude Haiku 5.5 at a 1M-token context window with up to 128K output tokens on the synchronous Messages API (300K on the batch API with a beta header). Google's own Gemini 3.8 Flash model page and DeepMind's model card both list an input window up to 1,048,576 tokens with a 64K-token output ceiling — half of Claude's.

For workloads that return long structured documents, transcripts, or large batches of generated records in a single call, that gap can matter more than the per-token price: a task that needs more than 64K tokens of output has to be split into multiple Gemini calls but can run in one Claude Haiku 5.5 call, which changes the effective cost and latency math even before token pricing is applied.

Knowledge cutoff and release dates

Gemini 3.8 Flash launched first, on September 2, 2026 according to DeepMind's model card, with a training data cutoff the card lists as March 2026 (noting some domains may only be current to January 2025). Claude Haiku 5.5 followed about five weeks later, launching October 7, 2026 per Anthropic's announcement, with a training data cutoff Anthropic's documentation lists as June 2026 — a few months more recent.

Speed and latency

Neither company publishes a standardized, independently measured latency benchmark for these two models against each other, so speed claims here come from each maker's own framing and from named customers quoted in each launch post — not from testing done for this article. Anthropic's announcement calls Claude Haiku 5.5 "our fastest model to date" and cites early adopters: Asana reports more than a 30% reduction in task-completion latency and up to 2.5x faster inference per agent turn, while Box reports roughly half the latency of Claude Haiku 4.5 in early testing. These are customer-reported figures from Anthropic's own post, not figures Anthropic or this site independently verified.

Google's DeepMind model card makes no comparable speed claim for Gemini 3.8 Flash; it instead notes the model can show "occasional slowness or timeout issues," without quantifying how often. Both models support an adjustable effort/thinking-depth setting that trades latency and token spend for answer quality, so real-world speed on either model depends heavily on how a developer sets that parameter for a given task.

Benchmark scores, by each maker's own tests

Both companies publish extensive benchmark tables, but almost none of the named tests line up exactly — different versions, different tool/no-tool settings, and different internal scaffolding make a clean side-by-side risky. Where the benchmark names are close enough to compare loosely, here is what each company reports about its own model:

Anthropic's own benchmark table also shows Claude Haiku 5.5 scoring 1578 on its "AA-Briefcase v1.1" knowledge-work test, 46.4% on its "FrontierCode 1.1 Main" coding test, and 46.4% on a visual-reasoning test it calls "Chartography" — all compared internally against Claude Haiku 4.5, Sonnet 5.5, and a rival model Anthropic refers to as "GPT-6 Luna." Google's model card separately reports Gemini 3.8 Flash at 73.7% on "DeepSWE v1.1" (long-horizon software engineering), 89.4% on "Terminal-bench 2.1," 61.4% on "Vals Finance Agent v2," and 86.2% on "CharXiv Reasoning," each compared against its predecessor, Gemini 3.7 Flash. Only the Terminal-Bench 4.0 result in the table above is a benchmark both companies name identically; Anthropic's Haiku 5.5 score there (39.2%) is more than double Google's reported Gemini 3.8 Flash score (19.1%), but because the two companies run their own harnesses, that gap should be read as directional rather than exact.

Multimodal input: what each model can actually read

This is where the two models diverge on capability rather than price. Anthropic's current model line, including Claude Haiku 5.5, is documented as supporting text and image input with text-only output. Google's Gemini 3.8 Flash model page lists a wider input set: text, images, video, audio, and PDF documents, still with text-only output. For teams whose "cheap model" job involves transcribing a call, summarizing a video, or parsing a native audio stream, Gemini 3.8 Flash's input range covers cases Claude Haiku 5.5 doesn't address natively — those teams would need to pre-process audio or video into text or images before sending it to Claude.

Where Claude Haiku 5.5 wins

On list price alone, Claude Haiku 5.5 is the cheaper of the two for any task measured mostly in text tokens: high-volume classification, routing, extraction, live chat support, and summarization, exactly the use cases Anthropic names in its own launch post. The longer 128K output ceiling also suits jobs that generate large structured outputs — bulk data transformation, long-form summarization, or generating many records in one call — without needing to paginate across requests the way a 64K-output model would. Anthropic's own further comparison of how Haiku 5.5 stacks up against its own larger siblings is a useful next read for teams deciding whether to stay cheap or step up within the Claude line itself.

Where Gemini 3.8 Flash wins

Gemini 3.8 Flash's case rests on modality and tool breadth rather than price: native audio, video, and PDF input means a single call can handle multimodal agentic work — watching a screen recording, reading a scanned document, or processing a voice note — that would otherwise require a separate transcription or vision step ahead of a text-only model. Google's own benchmark table also shows larger model-over-model gains on agentic computer-use and finance-agent tests, areas it's explicitly positioning the model for. Readers who want the fuller rundown of Gemini 3.8 Flash on its own terms can see Pandromeda's release-date and pricing explainer for the model.

What the sticker price doesn't show

Three caveats matter more than the headline numbers. First, Gemini 3.8 Flash's quoted rate is promotional and set to double on January 1, 2027 per Google's own pricing page; any cost comparison that ignores this will understate Gemini's 2027 cost by 2x. Second, Claude Haiku 5.5's pricing is tiered by prompt length, so a workload that regularly exceeds 100,000 tokens per call is effectively using a different, five-times-more-expensive rate card than one that stays under it — a distinction that doesn't show up in a simple $0.10/$0.50 headline. Third, neither company's benchmark table was produced by a neutral third party running both models under identical conditions; every score in this piece is self-reported by the maker of that model, labeled as such, and should be weighed accordingly. For the fuller numbers behind Claude Haiku 5.5 specifically, including Anthropic's full launch-day benchmark set, see Pandromeda's dedicated release coverage.

What's next

Google's own pricing schedule guarantees one concrete change: Gemini 3.8 Flash's per-token cost rises on January 1, 2027, which will narrow or widen the price gap with Claude Haiku 5.5 depending on whether Anthropic adjusts its own rates before then. Both companies ship small-model updates on roughly two-to-six-month cycles, so a Gemini 3.9 Flash or Claude Haiku 5.6-class release inside that window wouldn't be a surprise. Until then, the practical choice comes down to what a given pipeline actually sends: predominantly text workloads at high volume currently cost less on Claude Haiku 5.5, while workloads that need native audio, video, or PDF input in a single call currently have no Claude equivalent and point toward Gemini 3.8 Flash.

Written byTheo Park

Theo Park runs the AI desk at Pandromeda. He follows model launches from the frontier labs and the open-weight community, tracks the assistants and developer tools built on them, and explains what each release changes on pricing, capability and safety. His reporting leans on primary sources: model cards, technical reports, API documentation and the companies' own announcements.

More from Theo ParkAll AI coverage