The launch price was never the price On September 2, 2026, Google shipped Gemini 3.8 Flash at $0.75 per million input tokens and $3.75 per million output tokens, the same headline rate 3.7 Flash has run at all year. Buried in the same rate card is the part that doesn't show up in a benchmark chart: both numbers double on January 1, 2027, to $1.50 and $7.50. Source: Digital Applied, "Gemini 3.8 Flash Costs the Same Until It Doubles in January" Today's rate is a promo, not a price. Source: eesel AI, "Gemini 3.8 Flash pricing 2026: every rate, tier, and hidden cost" That alone would be a four-month-old story by the time most teams notice it on an invoice. What makes it worth stopping on now is the second number nobody put in the announcement: independent measurement already shows 3.8 Flash burning roughly 30% more output tokens per task than 3.7 Flash did, at the identical per-token rate, for a task that already costs about 40% more today than the same task did on the model it replaced. Source: Digital Applied, "Gemini 3.8 Flash Costs the Same Until It Doubles in January" Stack that on top of a rate that doubles in January, and a workload that looks flat on the rate card is heading toward roughly 2.8x its current cost, with no line item that says so until the invoice arrives. Gemini 3.8 Flash: standard, batch/flex, and priority tiers all double on January 1, 2027, four months after the September 2 launch at intro pricing What actually moves, tier by tier Every published tier scales together, so the doubling isn't limited to the headline number a rate-card screenshot usually shows: | Tier | Through Dec 31, 2026 | From Jan 1, 2027 | |---|---|---| | Standard (input / output) | $0.75 / $3.75 per M | $1.50 / $7.50 per M | | Batch / Flex (input / output) | $0.375 / $1.875 per M | $0.75 / $3.75 per M | | Priority, 1.8x (input / output) | $1.35 / $6.75 per M | $2.70 / $13.50 per M | Source: eesel AI, "Gemini 3.8 Flash pricing 2026: every rate, tier, and hidden cost" Output pricing already includes thinking tokens, so a model built to reason more before answering bills for that reasoning at the output rate, not some separate discounted line. Source: eesel AI, "Gemini 3.8 Flash pricing 2026: every rate, tier, and hidden cost" Cached input reads stay at $0.075 per million, the same 90% discount off standard input the tier has always carried, though neither source confirms whether that cache rate also doubles on January 1 alongside everything else. If it does, a team leaning on prompt caching to survive the doubling gets less cover than the sticker suggests. The 30% Google's own explanation predicts, but doesn't price Google's stated reason for 3.8 Flash's higher per-task cost isn't a pricing decision at all, it's a description of the model working as designed: it "works harder. On complex tasks, it exhibits greater diligence, executing extra reasoning steps, and calling tools iteratively." Source: Digital Applied, "Gemini 3.8 Flash Costs the Same Until It Doubles in January" That's a real capability upgrade, and on hard tasks it's probably the right trade. It's also, measured independently, about 30% more output tokens per task than 3.7 Flash at the same per-token price, which compounds to roughly 40% higher cost per task today, before the January change touches a single number on the rate card. Source: Digital Applied, "Gemini 3.8 Flash Costs the Same Until It Doubles in January" Run both multipliers on the same task, indexed to what it costs on 3.7 Flash today: 3.8 Flash costs about 1.40x that same task right now, on a rate card that hasn't changed since June. On January 1, the rate itself doubles on top of that already-inflated token count, landing the same task at roughly 2.80x what it cost on 3.7 Flash, an 180% increase, with zero change to the workload, the prompt, or the traffic volume in between. Same task, indexed to Gemini 3.7 Flash's cost today: 1.00x on 3.7 Flash, 1.40x on 3.8 Flash today from ~30% more output tokens, 2.80x on 3.8 Flash after the January 1, 2027 rate doubling stacks on top One agency-specific number makes the stakes concrete: a workflow costing $500 a month today could cost roughly twice that after the change, a real problem for any shop billing AI-assisted work at a fixed monthly rate quoted before September. Source: Enterprise DNA, "Google locks Gemini 3.8 Flash pricing through year-end, then doubles it" A fixed-price contract signed against September's rate card is already underwater against January's, four months before the invoice proves it. This is the third calendar-shaped cliff of 2026, not the first Google isn't inventing the mechanic. Claude Sonnet 5 launched June 30, 2026 at $2/$10 per million tokens, a third off the $3/$15 every Sonnet before it had charged, on a new tokenizer Anthropic's own docs say produces roughly 30% more tokens for the same text. The discount and the tokenizer inflation were designed to roughly cancel out, and that discount expired August 31, 2026. Different mechanic, same shape: an introductory number that reads as the price, and an expiration date that turns it into a preview. The other two cliffs this blog has tracked in 2026 aren't calendar-based at all, they're usage-based, and they're worth knowing about because they can stack with this one. Grok 4.6 doubles input and raises output 50% the instant a single request crosses 200,000 tokens, billed on the whole request, not just the tokens past the line. Gemini 3.1 Pro and GPT-5.6 Sol run near-identical mechanics at 200K and 272K tokens respectively. A team running a long agentic session on a Gemini model in January 2027 could, in principle, clear a context-length cliff and a calendar cliff in the same request. Meanwhile, frontier-tier API pricing overall rose 36.4% year over year through August 2026, while mid-tier and budget pricing fell 35.8% over the identical window, a 72-point spread that makes "the frontier default is safest" an increasingly expensive assumption to leave unreviewed. None of these labs are colluding on timing. What they share is a business incentive: launch at a number that wins the benchmark comparison and the migration decision, then let the rate card catch up to the real cost of serving the model once the switching cost has already been paid by the customer. Four steps before January 1 Pull your actual 3.8 Flash token counts, not the rate card. The 30% output-token increase is an average from independent measurement, not a guarantee for your workload. A reasoning-heavy agent loop could see more; a short-answer classification task could see less. Measure your own traffic before budgeting off someone else's number. Don't budget 2027 off September's invoice. A $500/month workflow today is a roughly $1,000/month workflow in January on rate alone, before any growth in traffic. Model the doubled rate against your actual current volume now, while there's still time to act on the number. Check whether your fixed-price client contracts assume September's Gemini pricing. If a consultancy or agency quoted a retainer against 3.8 Flash's intro rate, that quote is already stale relative to what the work will cost to deliver from January onward. Route on a quality floor instead of a model name, so a single vendor's calendar doesn't set your budget. A request pinned to "Gemini 3.8 Flash" pays whatever Google's rate card says on the day it runs. A request routed against a quality floor moves to whichever model currently clears that floor for the least cost, on that call, whether the cheapest option that week is a Gemini tier, a Claude tier, or something else entirely. import openai client = openai.OpenAI( base_url="https://api.getnadir.com/v1", api_key="YOUR_NADIR_KEY", ) response = client.chat.completions.create( model="auto", messages=conversation_history, No model name hardcoded, so a single vendor's January 1 rate card doesn't become your January 1 budget. The router scores this request against your quality floor on every call and picks the cheapest model that still clears it -- including away from Gemini 3.8 Flash the moment a cheaper option clears the same bar. ) print(response.model) # which model actually answered this call metadata = response.model_extra["nadir_metadata"] print(metadata["cost"]["total_cost_usd"]) The number that matters isn't $1.50 It's whichever number is cheapest for a given request on the day that request actually runs, because three different labs have now redrawn that number at least once each in 2026, on three different mechanics, with zero shared warning system between them. A team that priced its 2027 budget off a September rate card is betting against a pattern this blog has now documented three times in one year. Start free and let a quality-floor router absorb the next calendar cliff automatically, or read the complete guide to cutting LLM API costs for the rest of the stack that keeps working no matter which vendor moves its price next. FAQ Q: What exactly changes for Gemini 3.8 Flash pricing on January 1, 2027? A: Every published tier doubles. Standard pricing moves from $0.75/$3.75 per million input/output tokens to $1.50/$7.50. Batch and Flex pricing moves from $0.375/$1.875 to $0.75/$3.75. Priority tier pricing moves from $1.35/$6.75 to $2.70/$13.50. Google announced the change alongside the model's September 2, 2026 launch, with the intro rate holding through December 31, 2026. Q: Why does Gemini 3.8 Flash already cost more per task than 3.7 Flash, even before the price doubles? A: Google describes 3.8 Flash as working harder on complex tasks, running extra reasoning steps and calling tools iteratively. Independent measurement found that behavior burns roughly 30% more output tokens per task than 3.7 Flash at the identical per-token rate, which compounds to about 40% higher cost per task today, entirely separate from the January rate change. Q: Does the January 1, 2027 doubling apply to Gemini 3.6 Flash and 3.7 Flash too, or only 3.8 Flash? A: Published sources confirm the doubling for Gemini 3.8 Flash specifically. 3.7 Flash currently carries identical pricing to 3.8 Flash, but no source reviewed confirms whether 3.7 Flash's rate also changes on the same date. Treat 3.8 Flash as the confirmed case and verify 3.6 and 3.7 Flash directly against Google's published rate card before budgeting either one. Q: Has a major AI lab done this kind of calendar-based pricing change before in 2026? A: Yes. Claude Sonnet 5 launched June 30, 2026 at a third off Sonnet's standing rate, designed to roughly offset a new tokenizer that produces about 30% more tokens for the same text, and that discount expired August 31, 2026. Different mechanic, same pattern: an introductory rate that reads as the price until a fixed date turns it back into the standard one. Q: How should a team protect its 2027 AI budget from a change like this? A: Measure your own token usage on the model in question rather than trusting an industry-average estimate, model the post-doubling cost against current volume now, and check any fixed-price client work quoted against September's rate card. Structurally, routing requests against a quality floor rather than a hardcoded model name means a single vendor's calendar no longer sets your budget on its own, since the router can move traffic to whichever model clears the bar for less the moment the cheaper option changes.