Gemini 3.6 Flash vs GPT-5.6 vs Claude Haiku (2026)

TL;DR

• Fastest: Gemini 3.6 Flash, and it's not close. Around 300 tokens per second versus 100 to 150 for GPT-5.6 Luna.
• Cheapest capable coder: GPT-5.6 Luna at $1 / $6 per million tokens, and it out-scores Flash on coding benchmarks.
• Biggest context and multimodal: Gemini 3.6 Flash, 1M tokens with native video, audio, and PDF. Claude Haiku 4.5 tops out at 200K.
• Rock-bottom price: Gemini 3.5 Flash-Lite at $0.30 / $2.50, the sibling that launched the same day for high-volume simple work.
• No single winner: Flash for speed and context, Luna for cheap coding, Haiku for Anthropic-native text, Flash-Lite for the cheapest bulk jobs.

Google shipped Gemini 3.6 Flash on July 21, 2026, and the obvious question is how it lands against the other cheap, fast workhorses you'd actually run in production. Here's the honest side-by-side on price, speed, coding, and context, with a clear pick for each kind of job.

What actually launched?

Google dropped three models at once, not one. Gemini 3.6 Flash is the headline, a faster and cheaper update to 3.5 Flash. Alongside it came Gemini 3.5 Flash-Lite, a rock-bottom-priced tier for high-volume work, and Gemini 3.5 Flash Cyber, a restricted security specialist you can't just pick up in the API. For this comparison, Flash and Flash-Lite are the two that matter, since Cyber isn't generally available.

The models Flash competes with didn't stand still either. OpenAI's GPT-5.6 family arrived earlier in July with three tiers of its own, and its budget tier, Luna, is the direct rival here. Anthropic's Claude Haiku 4.5 rounds out the cheap-and-fast bracket. So the real fight is a four-way one, and it's more interesting than a plain "is the new Google model good" story.

Gemini 3.6 Flash vs GPT-5.6 Luna vs Claude Haiku 4.5: the numbers

Here's the head-to-head on the specs that decide a production choice. Prices are per million tokens.

The pattern jumps out fast. Gemini 3.6 Flash isn't the cheapest here. Luna and Haiku both undercut it on input and output, and Flash-Lite makes all three look pricey. What Flash buys you for the extra cost is speed and the widest context, which matters more for some jobs than others.

Which one is actually cheapest?

On raw sticker price, Gemini 3.5 Flash-Lite wins by a mile at $0.30 input and $2.50 output. But among the three main workhorses, GPT-5.6 Luna and Claude Haiku 4.5 both beat Gemini 3.6 Flash. Haiku is cheapest on output at $5, Luna sits at $6, and Flash is the priciest of the trio at $7.50.

Output price is where it stings. Why? Most real workloads generate far more output tokens than they take in, and a chatbot or an agent that writes long responses feels Flash's $7.50 rate on every single call. So if your task is output-heavy and doesn't need Flash's speed, Luna or Haiku saves real money at scale.

One more wrinkle. Luna ships with a 90% discount on cached input reads, which cuts cost further for repeat-context workloads like RAG. If you push the same system prompt or document set thousands of times, that cache discount can matter more than the sticker gap does.

What does each cost at scale?

Sticker prices feel abstract until you run the math. Take a modest workload of 1M input and 1M output tokens a day, and the monthly bills spread out fast:

• Gemini 3.6 Flash: about $270 a month
• GPT-5.6 Luna: about $210 a month
• Claude Haiku 4.5: about $180 a month
• Gemini 3.5 Flash-Lite: about $84 a month

That's Flash costing roughly 50% more than Haiku and over 3x Flash-Lite for the very same token volume. And most production traffic skews output-heavy, so the real gap runs wider than this even split shows. Is Flash's speed worth a 50% premium over Haiku? For a live chat UI where users watch each token arrive, often yes. For an overnight batch job nobody's staring at, almost never. That single question, latency-sensitive or not, decides more of the cost argument than the price sheet does.

Who's fastest?

Gemini 3.6 Flash, by a wide margin. It streams around 300 tokens per second, roughly double GPT-5.6 Luna's 100 to 150. For anything a user waits on, whether live chat, autocomplete, or an agent that has to feel responsive, that gap is the difference between snappy and sluggish.

Speed is Flash's real pitch. Google says 3.6 Flash also uses 17% fewer output tokens than 3.5 Flash to finish the same task, and takes fewer reasoning steps on multi-step workflows. Fewer tokens at a higher per-token rate can still net out cheaper on some jobs, so the sticker price alone undersells it for latency-sensitive, high-throughput use.

Does the 1M context gap matter?

Only if your workload actually fills it. Gemini 3.6 Flash, Luna, and Flash-Lite all carry a 1M-token window, while Claude Haiku 4.5 stops at 200K. Is 5x more context a real edge? For most chat and short-task work, not really. You'll never come close to 200K in the first place. It flips the moment you feed whole codebases, long PDFs, or hours of transcript into a single call, though. There, Haiku forces you to chunk your input and stitch the pieces back together, which adds latency and failure points, while the 1M models just swallow it whole. Big inputs? Then that gap alone can decide the choice, ahead of price or speed.

Where does each model win on quality?

They split the wins, which is why there's no clean overall champion. Each has a lane where it clearly leads:

• Coding goes to Luna. GPT-5.6 Luna out-scores Gemini 3.6 Flash on production coding benchmarks like SWE-Bench Pro and Terminal-Bench, and it's cheaper. For a coding agent on a budget, Luna is the value pick, with Claude Haiku 4.5 close behind on text-heavy coding inside its 200K window.
• Long context and multimodal go to Gemini. Flash wins long-context retrieval and chart-based reasoning by a margin the benchmarks call not close, and it takes native video, audio, and PDF. Its 1M window dwarfs Haiku's 200K.
• Hard knowledge work leans elsewhere. Flash gives up ground to bigger models like Claude Sonnet 5 on deep reasoning, so if quality on the hardest tasks matters more than speed or price, none of these cheap tiers is your answer.

Claude Haiku 4.5 is the awkward one here. It's strong on coding for its size, but at $1 / $5 it now trails Gemini 3.5 Flash-Lite and cheaper rivals on many benchmarks while costing more than Flash-Lite. It makes the most sense when you're already in the Anthropic ecosystem and want a fast Claude.

What changed from Gemini 3.5 Flash?

If you're already on 3.5 Flash, the upgrade is easy to justify. You get the same 1M context, a lower output price, and a smarter model. Output dropped from $9 to $7.50 per million, a 17% cut, while the model uses 17% fewer output tokens for the same work. On output-heavy jobs that's a double saving.

You also get a knowledge cutoff that finally moved from January 2025 to March 2026, built-in Computer Use for agent tasks, and better tool-calling efficiency. For most existing 3.5 Flash workloads, moving to 3.6 Flash is cheaper and better with no real downside. That combination is rare enough to just do it.

And what about Gemini 3.5 Flash Cyber?

It's the odd one out, and you probably can't use it. Gemini 3.5 Flash Cyber is a restricted security specialist, tuned for cyber-defense and threat-analysis work, and it isn't part of the open Gemini API the way Flash and Flash-Lite are. Access is gated. For the vast majority of developers building normal apps, it simply isn't on the table, which is why this comparison sticks to the two public Gemini models. Worth knowing it exists, mostly so you're not hunting for an API endpoint that will never answer your key.

Which should you use?

There's no single right answer, and that's the honest takeaway of the whole comparison. Match the model to the job in front of you, not to whichever launched most recently:

• Pick Gemini 3.6 Flash for latency-sensitive apps, long-context or multimodal work, and agents that need speed. It's the fastest and has the widest usable context.
• Pick GPT-5.6 Luna for cost-sensitive coding and high-volume tasks where its cache discount and coding edge pay off. It's the cheapest capable coder of the group.
• Pick Claude Haiku 4.5 when you're Anthropic-native and want fast Claude for text and coding inside a 200K context.
• Pick Gemini 3.5 Flash-Lite for the highest-volume, simplest jobs, where price per token beats everything else.

The catch with a multi-model strategy is the plumbing. Three vendors means three accounts, three SDKs, and three billing setups. Routing all of them through one OpenAI-compatible endpoint like Velokey skips that, so you can send coding calls to Luna and fast multimodal ones to Flash on a single key. For deeper dives, our GPT-5.6 Sol vs Terra vs Luna breakdown covers OpenAI's tiers, the GLM vs GPT vs Claude comparison maps the frontier tier, Claude Sonnet 5 is the step up for knowledge work, and Kimi K3 is another cheap frontier option worth a look.

Frequently Asked Questions

Is Gemini 3.6 Flash cheaper than GPT-5.6?

Not than every tier. Gemini 3.6 Flash costs $1.50 input / $7.50 output per million tokens, which is more than GPT-5.6 Luna ($1 / $6) but far less than GPT-5.6 Terra ($2.50 / $15) or Sol ($5 / $30). Against Luna specifically, Flash is the pricier but much faster option.

Is Gemini 3.6 Flash better than Claude Haiku 4.5?

It depends on the job. Flash has 5x the context (1M vs 200K), stronger multimodal, and higher speed. Haiku 4.5 is cheaper on output and strong on coding for its size. For long-context or multimodal work, Flash wins; for Anthropic-native text and coding in a smaller context, Haiku holds up.

What's the difference between Gemini 3.6 Flash and 3.5 Flash-Lite?

Flash is the faster, more capable model; Flash-Lite is the rock-bottom-priced one. Flash costs $1.50 / $7.50 and leads on speed and reasoning. Flash-Lite costs $0.30 / $2.50 and targets high-volume, simpler tasks where cost per token matters most. Both keep the 1M context window.

Which cheap model is best for coding?

GPT-5.6 Luna, on current benchmarks. It out-scores Gemini 3.6 Flash on production coding tests like SWE-Bench Pro and Terminal-Bench while costing less. Claude Haiku 4.5 is a solid second for text-heavy coding within 200K context. Gemini 3.6 Flash trails both on pure coding.

Should I upgrade from Gemini 3.5 Flash to 3.6 Flash?

Yes, for most workloads. 3.6 Flash cuts output price from $9 to $7.50, uses 17% fewer output tokens for the same task, moves the knowledge cutoff to March 2026, and adds built-in Computer Use. Same 1M context, lower cost, better results, so there's little reason to stay on 3.5.

Where can I access Gemini 3.6 Flash?

It's live in the Gemini API through Google AI Studio and Android Studio, and in Google Antigravity. You can also reach it, GPT-5.6, and Claude Haiku through a single OpenAI-compatible gateway if you'd rather not juggle three separate provider accounts and keys for a multi-model setup.

Does Gemini 3.6 Flash support tool use and Computer Use?

Yes. Gemini 3.6 Flash ships with native tool calling and built-in Computer Use, so it can drive browser and desktop actions for agent workflows out of the box. Google also reports it uses fewer tool calls and reasoning steps than 3.5 Flash to finish the same multi-step task, which trims both latency and token cost on agentic jobs.