On this page · 9 sections
Summary. As of July 2026, the budget tier of frontier-class models has three main contenders. Google's Gemini 3.5 Flash-Lite, launched July 21, 2026, costs $0.30 per million input tokens and $2.50 per million output. OpenAI's GPT-5.6 Luna, part of the family that reached general availability on July 9, 2026, runs $1 input and $6 output. Anthropic's Claude Sonnet 5, out since June 30, 2026, sits at introductory rates of $2 input and $10 output through August 31, 2026, then $3 and $15. On a representative input-heavy workload of 10 million input and 2 million output tokens per day, that spread works out to roughly $240, $660 and $1,200 per month respectively. The sticker price, though, is not the number that lands on your invoice: tokenizer differences, output-to-input ratio, cached-input discounts and quality-driven retries move the real cost by a wide margin. This is the cost math for teams choosing a model for high-volume work.
The reason this tier matters more than the flagships is volume. A support agent, a document pipeline or a retrieval system runs the same prompt shape millions of times, so a few cents per million tokens compounds into the largest line on the AI bill. Picking the right budget model is a bigger lever than squeezing the flagship. Our flagship model comparison of Gemini 3.5 Pro, GPT-5.6 and Claude Fable 5 covers the top of the range; this piece is about the tier you actually run at scale.
The three budget models at a glance
Every rate below is the vendor's published July 2026 API price per million tokens.
| Model | Input ($/1M) | Output ($/1M) | Launch | Cached input |
|---|---|---|---|---|
| Gemini 3.5 Flash-Lite | $0.30 | $2.50 | July 21, 2026 | Discounted (see Google pricing) |
| GPT-5.6 Luna | $1.00 | $6.00 | July 9, 2026 | 90% off input on cache reads |
| Claude Sonnet 5 (intro) | $2.00 | $10.00 | June 30, 2026 | Up to 90% off cached input |
| Claude Sonnet 5 (standard, from Sept 1) | $3.00 | $15.00 | June 30, 2026 | Up to 90% off cached input |
| Gemini 2.5 Flash-Lite (legacy floor, retiring Oct 16, 2026) | $0.10 | $0.40 | Earlier | Discounted |
Two things stand out. Flash-Lite's output rate ($2.50) undercuts Luna's ($6) by more than half and Sonnet 5's ($10 intro) by three-quarters. And the legacy Gemini 2.5 Flash-Lite is still the absolute floor at $0.10/$0.40, but Google retires it on October 16, 2026, so it is not a floor to build a 2027 roadmap on.
These three are not the same class of model. Anthropic positions Sonnet 5 as its most agentic Sonnet, with performance close to the more expensive Opus 4.8 (priced at $5/$25) at a lower price. OpenAI calls Luna its "fastest and most affordable" GPT-5.6 tier and reports that Luna outperforms the older Opus 4.8 on its coding-agent index at roughly one-quarter the estimated cost. Flash-Lite is the smallest and cheapest of Google's current generation, built for high-volume, latency-sensitive work rather than the hardest reasoning. So price alone does not decide this; it sets up the real question, which is cost per acceptable answer.
The real cost math
Take a representative high-volume workload: 10 million input tokens and 2 million output tokens per day. That input-heavy shape is typical of retrieval-augmented generation and support triage, where the model reads a lot of context and writes a short answer. At published July 2026 rates:
| Model | Daily cost | Monthly (30 days) | Relative to Flash-Lite |
|---|---|---|---|
| Gemini 3.5 Flash-Lite | $8.00 | ~$240 | 1.0x |
| GPT-5.6 Luna | $22.00 | ~$660 | 2.75x |
| Claude Sonnet 5 (intro) | $40.00 | ~$1,200 | 5.0x |
| Claude Sonnet 5 (standard) | $60.00 | ~$1,800 | 7.5x |
Now flip the workload to output-heavy, 2 million input and 10 million output tokens per day, the shape of content generation or long-form drafting. Because output tokens cost five to eight times what input tokens cost across all three vendors, the bill changes character: Flash-Lite lands near $25.60 per day, Luna near $62, and Sonnet 5 (intro) near $104. The lesson is that your input-to-output ratio, not just the model, decides which rate dominates your invoice. Measure your actual ratio before you pick.
These figures are illustrative arithmetic on the sourced rates, not vendor quotes. They exist to show the shape of the decision, and the shape is stark: on an input-heavy pattern Flash-Lite is roughly one-third the cost of Luna and one-fifth to one-seventh the cost of Sonnet 5.
Why the sticker price is not your real cost
Four factors move the true number away from the rate card. Miss them and a "cheap" model quietly becomes the expensive one.
First, tokenizers differ. The same paragraph does not become the same number of tokens across vendors. Anthropic states plainly that Sonnet 5 uses an updated tokenizer under which the same input can map to roughly 1.0 to 1.35 times more tokens than before, depending on content type, and it set the introductory price so the move from Sonnet 4.6 is roughly cost-neutral. At the top of that range, Sonnet 5's effective cost on the input-heavy workload above is closer to $48 per day than $40. Always compare cost on your own text, tokenized by each provider, not on the headline rate.
Second, caching. All three vendors discount repeated context heavily. OpenAI gives GPT-5.6 a 90% discount on cache reads, bills cache writes at 1.25 times the uncached input rate, and holds a 30-minute minimum cache life. Claude offers up to 90% savings on cached input. If your prompts share a large fixed prefix, a system prompt, a policy document, a schema, the cached-input rate, not the standard rate, is your real input cost, and that can swing the ranking.
Third, the output-to-input ratio, covered above: output is the expensive side of the meter everywhere.
Fourth, and largest, quality-driven retries. If a cheaper model clears your quality bar only 80% of the time and you retry or escalate the other 20%, you pay for the cheap call, the retry, and sometimes an expensive-model fallback. A model that is half the price but needs two attempts is not half the price. The honest budget-model question is not "what is the rate" but "what does one accepted answer cost, including the failures". We walk through that routing logic in our LLM hybrid-routing decision framework.
When each model wins
The pricing points to a clear division of labour rather than one winner.
Choose Gemini 3.5 Flash-Lite when volume dominates and each task is bounded: classification, extraction, routing, short retrieval answers, tagging, moderation triage. At $0.30/$2.50 it is the cheapest of the three current-generation options by a wide margin, and for well-scoped tasks a smaller model often clears the bar. Its risk is complex, multi-step reasoning, where a cheaper model can fail more often and erase the saving through retries.
Choose GPT-5.6 Luna when you want budget pricing but need stronger agentic behaviour, tool use and coding than a lite model gives. OpenAI reports Luna scoring 51.2 on the Artificial Analysis Intelligence Index v4.1 and 74.6 on its Coding Agent Index v1.1, and beating the older Opus 4.8 on coding at about a quarter of the estimated cost. At $1/$6 it is the middle option here on both price and capability. Simon Last, co-founder at Notion, put the tier's value directly: "Terra and Luna also punch well above their price. Many agents running GPT-5.5 perform just as well on Terra for half the cost and 16% fewer tokens."
Choose Claude Sonnet 5 when the work is genuinely agentic, long-horizon coding, multi-step tool use, brownfield debugging, and quality and follow-through matter more than raw cost. Anthropic positions it near Opus 4.8 at a lower price, with adjustable effort levels so you can trade cost against depth per task. It is the most expensive of the three, and for high-stakes automation that is often the right trade, because a wrong action is dearer than a few dollars of tokens. For teams already on the Opus tier, our note on cutting token spend with Claude effort levels applies to Sonnet 5's effort control too.
Capability, not just price
Cheapest on the rate card is not cheapest per accepted answer, so weigh capability signals alongside cost. On OpenAI's own published comparison, GPT-5.6 Luna reaches 92.3% on GPQA Diamond, 83.3% on BrowseComp and 50.3% on Agents' Last Exam, close behind the flagship Sol on several axes at a fraction of the price. Anthropic's charts show Sonnet 5 covering a wider cost-performance range than its predecessor Sonnet 4.6 and matching Opus 4.8 on some tasks at higher effort. Google's Flash-Lite trades peak reasoning for throughput and latency. The buying rule that follows: benchmark the two or three finalists on a sample of your own tasks, score acceptance rate, and divide your real monthly spend by accepted answers. That number, not the headline rate, is the one to compare. The token-efficiency angle, how many tokens a model burns to reach the same answer, is covered in our piece on Gemini 3.6 Flash token efficiency and agent cost, and it often matters as much as the per-token price.
India-specific considerations
For teams building in India, three points follow.
The unit that hits the ledger is still US dollars, because all three are US-hosted APIs billed in USD. At the input-heavy volume above, the monthly gap between Flash-Lite (~$240) and Sonnet 5 standard (~$1,800) converts to a large rupee difference at any exchange rate, so for cost-sensitive Indian SaaS and D2C products the budget tier is not a nice-to-have, it is the difference between a viable unit economic and a broken one. Convert every figure at your live USD-INR rate before you model margins.
Data residency and the Digital Personal Data Protection Act, 2023 (DPDP) shape more than price. These APIs process prompts on US infrastructure by default, so where prompts carry personal data, the design question is what you send, how you minimise and redact it, and whether a regional or self-hosted path is needed. We design applications aligned with DPDP requirements rather than claiming certification, and for the strictest cases an open-weights model on your own infrastructure can be both cheaper at very high volume and cleaner on residency, a path we cover in running local LLMs in production with vLLM, Ollama and LM Studio.
Finally, routing. The cheapest production setup is rarely one model. Send bounded, high-volume tasks to Flash-Lite, route the harder cases to Luna or Sonnet 5, and measure the blended cost per accepted answer. The same discipline sits behind our cloud FinOps guidance for Indian teams: measure first, then cut where the data points.
FAQ
How eCorpIT can help
eCorp Information Technologies (eCorpIT) is a Gurugram-based, senior-led engineering organisation, ISO 27001:2022 certified and assessed at CMMI Level 5, that helps teams pick and route budget-tier models without paying for it in quality. We benchmark the finalists on your own tasks, wire up caching and hybrid routing, and report cost per accepted answer instead of a raw token rate, so the model choice is grounded in your data and your DPDP constraints. If you are sizing or migrating an LLM workload, talk to our team about a cost and routing review, or read about our LLM migration and cost optimization service.
References
_Last updated: July 28, 2026._
