Gemini 3.8 Flash Thinking Tokens: Real Task Cost at $0.41

Gemini 3.8 Flash lists at $0.75 input and $3.75 output per million tokens through December 2026, identical to 3.7 Flash. The price nobody quotes is thinking tokens, billed at output rates with reasoning on by default at medium. Independent runs put real evaluated task cost at $0.24 low effort, $0.41 medium, and $0.58 high. I rebuilt our spend model around those numbers at SaaSNext, and the promo-to-regular doubling in January changes every annual forecast.

What finance needs in one paragraph:

• Response price equals visible output tokens plus hidden thinking tokens, all at $3.75 per million promo.
• Medium effort is the practical default at $0.41 per task with intelligence 57. High buys 59 for $0.58.
• Standard rates double January 1, 2027 to $1.50 input and $7.50 output. Model both halves of the year now.

I learned this the embarrassing way. Our August forecast priced Gemini work from visible output tokens only and came in 46 percent under actuals. The gap was thinking tokens on medium, exactly as the docs describe for anyone who reads the pricing footnotes. September forecast uses metered totals and landed within 6 percent.

The pricing table with the footnotes included

Google's standard tier is simple until thinking enters:

Two lines decide real spend. First, output price includes thinking tokens, exposed as total thought tokens in usage metadata. Second, current prices are promotional and double on January 1. A workload that costs $1,000 monthly in October costs $2,000 in January at identical usage. I state this in every proposal now because clients remember the promo and forget the doubling.

For the benchmark side of this model, my Claude Code versus Gemini coding truth covers where the intelligence goes, and Qwen versus Gemini audio bench covers the analyst workloads.

Production war story 1: the 46 percent forecast miss

Our support triage fleet ran Gemini 3.8 Flash at default settings through August. I estimated $620 monthly from average visible output of 1,900 tokens per ticket across 11,000 tickets. Actual bill was $905. Difference was thinking tokens averaging 1,650 per ticket, nearly doubling billed output. Same tickets, same quality, 46 percent more spend than my sheet.

Fix was measurement plus effort tuning. I pulled total thought tokens per ticket for a week, set low effort for classification steps and medium for resolution drafting, and capped max thinking on routine intents. September run rate is $590 with resolution quality flat. The only code change was passing thinking level per step instead of defaulting everything to medium.

Second finding from the same data: cache reads at $0.075 per million are the cheapest lever available. System prompts plus few-shot examples cached across tickets cut input spend 31 percent. Cache storage at $0.50 per million per hour sounds scary until you divide by thousands of hits. Our cache ROI is roughly 9 to 1.

Effort levels: what each step really buys

Independent measurements across low, medium, and high tell a clean story:

Low to medium is the trade worth making: plus 5 intelligence and nearly 5 agentic points for $0.17. Medium to high is the trade to question: plus 2 intelligence for $0.17 more, a 41 percent cost jump for a marginal gain. I default everything to medium, drop to low for classification and extraction, and reserve high for tasks where one extra success pays for a hundred failures.

Speed is genuinely elite at 299 tokens per second, rank 3 of 196 measured models. Time to first token regressed to 13.21 seconds from 12.01 on 3.7, about 10 percent slower to start. Async agents never notice. Interactive chat does. Route accordingly.

My coordinator fleet wiring for these effort choices lives in Claude Code Projects at 200 threads, where per-thread effort caps are the main spend throttle.

Production war story 2: the January doubling nobody modeled

A client signed a 12-month automation contract in late August priced from September promo rates with no escalation clause. January doubling turns their $2,400 monthly inference line into $4,800 overnight. Margin on the deal goes negative in Q1. I caught it reviewing the pricing page during this article's research, three weeks after signing.

Renegotiation was uncomfortable but successful. We moved the client to a two-tier clause: promo rates through December, regular rates from January, with a batch-API commitment that halves both. We also shifted overnight bulk jobs to batch at $0.375 and $1.875, cutting that slice 50 percent at both rate cards. Lesson is now policy: every proposal models both halves of the year, and every contract carries the January clause with batch options attached.

I also test front-end taste separately after finding 3.8 slightly behind 3.7 on design arena scores, 1311 Elo against 1318. For dashboard and marketing page generation, that half-step matters more than two intelligence points. Taste evals are cheap. Run them.

Cost control playbook that survived contact with billing

EFFORT_POLICY = {
    "classify": "low",      # extraction, routing, labels
    "draft": "medium",      # summaries, responses, code drafts
    "reason": "high",       # hard analysis, disputed cases only
}
BUDGET_PER_TASK = 0.60  # page when metered total exceeds this
CACHE_TTL_HOURS = 4     # system prompt plus examples

def check_spend(metered_input, metered_output, thinking):
    billed_output = metered_output + thinking
    cost = metered_input * 0.75 / 1e6 + billed_output * 3.75 / 1e6
    return cost

Four rules carry most of the savings. First, set thinking level per step, never globally. Second, read total thought tokens from metadata on every call and log it next to visible output. Third, cache system prompts with a 4-hour TTL and measure hit rates weekly. Fourth, move delay-tolerant bulk work to batch or flex at half rates. Combined effect on our fleet: 38 percent lower spend per resolved ticket with quality scores unchanged.

Batch economics deserve emphasis. Overnight backfills, eval suites, and report generation lose nothing to batch latency and gain 50 percent. Priority inference at $1.35 and $6.75 is the opposite lever, reserved for launch-day spikes where latency buys revenue. Most teams need neither lever most days. The ones that do should price them explicitly.

When NOT to optimize for token price

Skip effort downgrades on tasks where failure costs dwarf inference: refunds, access control, medical-adjacent triage. One wrong answer at $0.24 costs more than ten careful ones at $0.58. Skip batch for anything user-facing with latency promises. And skip provider switching on price alone when harness tuning, caching, and effort policy deliver larger savings with zero migration risk.

For durable job design around these cost controls, see LangGraph on Temporal for checkpointing that prevents paying twice for the same work after restarts.

Verification checklist for your spend model

1. Meter thinking tokens separately for one week before forecasting anything.
2. Price medium as default, low for classification, high only with written justification.
3. Model January 2027 regular rates in every annual forecast starting today.
4. Move bulk async work to batch and measure the 50 percent saving directly.
5. Review cache hit rates and effort mix monthly. Drift returns silently.

My verdict: $0.41 medium is honest value for 57 intelligence at 299 tokens per second, as long as you count thinking tokens and plan for January. The teams getting burned are pricing visible output and forgetting both footnotes.

By Deepak Bagada, Founder and Editor-in-Chief at Daily AI World. I own inference economics at SaaSNext and reconcile every forecast against metered bills. More at deepakbagada.in.