Gemini API Pricing 2026: Is Google Finally the Cheapest Frontier Model?

What Google charges in 2026

Google runs five billable Gemini families in July 2026 — gemini-3.1-pro, gemini-3-pro, gemini-3-flash, gemini-2.5-pro, and gemini-2.5-flash — plus a still-functional free tier inside Google AI Studio for evaluation and lightweight production. gemini-flash-lite was deprecated on June 1, 2026, leaving gemini-2.5-flash as the cheapest current model at $0.075/$0.30 per million tokens. The flagship Gemini 3.1 Pro charges $2/$12 per 1M tokens up to the 200K context threshold, doubling to $4/$18 beyond.

The pitch is simple: Google is the cheapest frontier model with the longest context window and the only production free tier remaining in 2026. Whether that holds depends on which tier you pick and which threshold you cross.

The full Gemini rate card (July 2026)

All rates per 1M tokens. The “≤200K context” column applies when total request tokens are under 200,000. The “>200K context” column kicks in past 200K. Image generation is billed by output tokens — one 1024×1024 image costs 1,290 tokens.

Image, audio, and embeddings

The 200K context tier vs long-context tier

Google’s tiered pricing is the most consequential pricing detail of 2026 and the easiest one to miss. The same model on the same call can double in price if your context crosses 200K tokens. Plan accordingly.

• ≤200K context tier: gemini-3.1-pro at $2/$12, gemini-3-flash at $0.50/$3, gemini-2.5-pro at $1.25/$10.
• >200K context tier: same models, rates double. gemini-3.1-pro jumps to $4/$18, gemini-3-flash to $1/$6, gemini-2.5-pro to $2.50/$15.
• Context caching: $0.20/M on 3.1 Pro, $0.05/M on 3 Flash, $0.02/M on 2.5 Flash. Caching discount is roughly 90% on every tier.

The right move on long-context Gemini calls is to summarize the long context down to ≤200K when possible, then push the full document through a cheaper model only when the summary isn’t enough. Doubling the rate for a single over-context call silently doubles a month’s worth of bills.

Free tier vs paid

Google is the last frontier vendor with a production free tier, and it’s still useful in 2026. The Google AI Studio free tier includes:

• 5 requests per minute on gemini-2.5-flash and gemini-3-flash.
• 100 requests per day across all models, including Pro tiers.
• Free image generation up to 100 images/day on Imagen 3.
• Free embeddings up to 1,000 requests/day on text-embedding-004.

Once you cross the daily cap, Google gives you the option to upgrade to a paid tier without disabling the project. The paid free-tier-bridge starts at $0.075/M input on Flash-Lite legacy and $0.075/M cached input on 2.5 Flash — both lower than OpenAI’s cheapest cached tier and roughly half of Anthropic’s Haiku rates.

If you’re a solo founder evaluating a new agent idea, Google AI Studio is the only place in 2026 where you can ship a working prototype for $0/month of API spend. OpenAI and Anthropic both cut their free tiers through 2025.

Three real workloads with actual bills

Workload A — 1M-token codebase analysis

You’re indexing an entire monorepo (1M tokens of code + dependencies) for question-answering. 100 queries per day, each query adds ~5K tokens of follow-up context.

• On gemini-3-flash with 1-hour cache: 1M cache write ≈ $0.05; per-query cost is ~5K × $0.50/M = $0.0025 input + 1K output × $3/M = $0.003. Daily: 100 × $0.0055 = $0.55; weekly: $3.85.
• On gemini-2.5-pro uncached: $1.25 × 1M × 100 = $125/day, $3,750/week. Don’t do this — caching is mandatory at this scale.
• On gemini-3.1-pro with cache: similar cached cost; output is 2.4x the cost ($12 vs $30 on Opus) but Opus wins on deep reasoning. Pick Pro unless you have a measured task that requires Opus-tier reasoning.

Workload B — 1,000-image batch

A marketing agency renders 1,000 product photos at 1024×1024 with prompt templating. Small text input (caption + brand context).

• On imagen-3 at $0.039/image: 1,000 × $0.039 = $39. Per-image roughly half of gpt-image-1.5 medium tier.
• On gpt-image-1.5 medium: ~$0.04/image = $40.
• On Imagen at 2048×2048 ($0.156/image): $156. Use only for hero images.

Workload C — 10-hour research task (deep agent)

A research agent ingests 80 web pages, runs 30 tool calls, and produces a 5K-token synthesis. Total tokens ~500K input (cache-eligible), 15K output (synthesis + summaries).

• On gemini-3.1-pro with caching (≤200K tier): cached input 80% at $0.20/M, fresh input 20% at $2/M; output 15K × $12 = $180. Input ~$52. Total $232 per research task.
• On gemini-3-flash with caching: cached input $0.05/M, fresh $0.50/M; output 15K × $3 = $45. Input $4.50. Total $50 per task. Quality is sufficient for short factual research; longer synthesis needs Pro.
• If the 500K input crosses the 200K threshold on a single task: rate jumps to $4/$18. Bill doubles to ~$390/task. Plan around the threshold or split the research.

When Gemini wins, when it loses

Use Gemini when:

• You need the cheapest frontier model on high-volume classification. gemini-2.5-flash at $0.075/M input is roughly 3x cheaper than gpt-5.4-nano and 13x cheaper than claude-haiku-4-5.
• You need a 1M-token context window on a budget. Gemini’s 1M tier is the only one in 2026 under $4/M input; Anthropic and OpenAI stop at 200K without long-context surcharge.
• You need built-in image gen via Imagen 3, video gen via Veo, and audio transcription in one vendor.
• You want a free tier that survives into production. Google AI Studio’s free tier is unmatched in 2026.
• You have EU/APAC data residency requirements and want local-region endpoints with no surcharge.

Avoid Gemini when:

• You’re running agentic tool-use loops with many tools. Claude 4.6 has the best tool-use adherence; Gemini hallucinates tool arguments more often at ≥6 tools.
• You depend on best-in-class reasoning for hard multi-step problems. Opus 4.7 and gpt-5.5 win on benchmarks; Gemini Pro is close but slightly behind.
• Your workflow needs reliable structured output with strict schemas. Gemini has improved but still produces more JSON drift than OpenAI on tight schemas.

The hidden costs that nobody quotes you

Rate cards look clean until you start shipping. Four things drive your real Gemini bill higher than the dashboard suggests, and they all bite on the same month.

Threshold overflow on long-context calls. A single 210K-token call on gemini-3.1-pro silently bills at $4/$18 instead of $2/$12. For solo operators running nightly research, this can double the bill for one over-threshold job. Audit your context sizes before pushing to production.

Free-tier rate limits during growth. Google AI Studio’s free tier gives 100 requests/day. The moment you push a tool that hits 105 requests, you start paying — and the dashboard charges the same Flash rates as if you’d been on paid the whole time. Cap your traffic explicitly with rate-limit middleware, or move to a paid project before launch.

Imagen retries. Imagen 3 has a ~3–5% failure rate on first-render attempts with structured prompts. A workflow with three retries per image triples your $0.039 budget to $0.117. Use conservative prompts and pre-validate caption length before queueing renders.

Veo (video) is shockingly cheap per second but very expensive per project. $0.40/sec at 1080p sounds cheap. A 30-second commercial is $12 of compute. A 100-video campaign is $1,200. Track per-project Veo spend the same way you’d track contractor invoices.

FAQ

How much does Gemini 3.1 Pro cost per million tokens?

$2.00 input / $12.00 output per 1M tokens for context under 200K. Past 200K, the rate doubles to $4/$18. Cached input is $0.20/M — a 90% discount on the under-200K rate.

Is the Google AI Studio free tier still available in 2026?

Yes. Free tier includes 5 requests per minute on Flash models, 100 requests per day across all models including Pro tiers, 100 free Imagen images per day, and 1,000 free embedding requests per day. It’s the only production free tier among the frontier vendors.

Was Flash-Lite deprecated?

Yes — gemini-flash-lite was deprecated on June 1, 2026. gemini-2.5-flash at $0.075/$0.30 is the new floor for the Gemini API.

How much does Imagen 3 cost per image?

1,290 output tokens = $0.039 per 1024×1024 image at the gemini-2.5-flash-tier pricing. Larger sizes scale linearly. Veo 3 video is $0.40/sec at 1080p resolution.

Does Gemini have prompt caching?

Yes — context caching on Gemini gives roughly 90% off input across all current models. Cache TTL is configurable; default is 30 minutes, extendable to 24 hours on explicit long-context caching.

Should I pick Gemini 3 Pro or 3.1 Pro?

Same price, same rates. 3.1 Pro is the GA model and the recommended pick for new deployments in July 2026. 3 Pro is kept on the dashboard for transitional customers but receives slower updates.

What is the best cheap Gemini model for high-volume classification?

gemini-2.5-flash at $0.075/$0.30 per 1M tokens is the cheapest current option and outperforms gpt-5.4-nano on most classification benchmarks while costing roughly 1/3 the price.

Does Gemini support fine-tuning in 2026?

Fine-tuning is available on gemini-2.5-flash and gemini-2.5-pro via the Gemini Tuning API. Pricing on tuned models stays the same as the base — you pay for input/output tokens only, no per-tuning-hour fee. The 3.x family added tuned-model support in June 2026 but is still on a gated beta.

How does Google charge for grounding (search)?

Grounding with Google Search is enabled via the google_search tool and is billed at $35 per 1,000 queries on top of input/output tokens. For research agents this is the cheapest grounding layer among the three frontier vendors — Anthropic charges $10 per 1,000 web searches with a stricter quota.

Verdict

Yes — Google is genuinely the cheapest frontier model in 2026 once you account for the tier that fits your workload. gemini-2.5-flash at $0.075/M is a third of gpt-5.4-nano, and gemini-3-flash at $0.50/$3 undercuts claude-haiku-4-5 for any task above a few hundred tokens of context. The two non-obvious traps: the >200K context tier that doubles your rate on a single over-threshold call, and the tool-use quality that lags Claude and OpenAI at high tool counts. Plan around those and Google becomes the highest-margin vendor in your stack — particularly if you’re running an agency on a 60% gross margin target.

Default routing for a multi-vendor agent in July 2026: Gemini Flash for classification and bulk ingestion, Sonnet 4.6 for tool-heavy agent loops, Opus 4.7 only on the synthesis step. That routing alone typically cuts a Claude-only bill by 40–60%. The per-image cost gap is also real — Imagen 3 at $0.039/image is the cheapest 1024×1024 render in the industry today, and Google is the only vendor that bundles free-tier image gen, video gen, and embeddings into the same free project you can keep running in production all month. The strategic play: ship the prototype on the free tier, graduate to paid Flash when the volume justifies it, and reserve Gemini Pro for the workloads that need the quality bump.

Updated July 3, 2026. Reviewed against the Google AI Studio and Gemini API dashboards. Next review: August 1, 2026.

More from