On this page · 12 sections
- Why this comparison changes on 1 January 2027
- Gemini 3.8 Flash: after 1 January
- Claude Haiku 5.5
- GPT-6 Luna
- Head-to-head comparison
- Published benchmarks, labelled by who ran them
- Worked cost example
- Decision table: which cheap model to pick
- Data residency: India, EU and US
- Migration checklist: moving traffic between these models
- How eCorpIT can help
- References
Summary. From 1 January 2027, Gemini 3.8 Flash costs $1.50 per million input tokens and $7.50 per million output tokens on Google's paid API, 15 times the $0.10 and $0.50 list price shared by Claude Haiku 5.5 (released 7 October 2026) and GPT-6 Luna (released 22 September 2026). All three offer context windows of about 1 million tokens, a 90% discount on cache reads and a 50% batch discount, but long prompts are priced differently. Haiku 5.5 bills the whole request at $0.50 and $2.50 once the prompt passes 100,000 tokens, Luna charges 2x input and 1.5x output above 272,000 input tokens, and Gemini stays flat. On independent benchmarks Haiku 5.5 leads: Artificial Analysis Intelligence Index v5.0 scores it 36, against 31 for Gemini 3.8 Flash and 28 for Luna, and the Vals AI index puts the three at 58.73%, 55.95% and 52.32%. Luna is cheapest per task on both, at $0.15 and $0.45, because it uses fewer tokens. Gemini keeps a real edge on native video and audio input, Google Search grounding and Google Cloud integration. Our verdict: default to Haiku 5.5 for agentic and knowledge work under 100,000 tokens, use Luna for high-volume short-output jobs and 100,000 to 272,000-token reads, and keep Gemini where multimodal input or grounding decides.
About this article. This comparison is for CTOs, platform engineers and product teams choosing a low-cost model for 2027 budgets. We checked prices, limits and release dates against Google's, Anthropic's and OpenAI's pricing pages, model pages and changelogs, and benchmark figures against the Artificial Analysis and Vals AI model pages, on 1 November 2026. eCorpIT ran no tests of its own. Vendors change prices often, so re-check before you commit; no vendor paid for inclusion. Researched with AI assistance and reviewed by our editorial team.
Why this comparison changes on 1 January 2027
Three launches in five weeks reset the budget tier. On 2 September Google released Gemini 3.8 Flash at an introductory $0.75 input and $3.75 output per million tokens, with a footnote that $1.50 and $7.50 apply from 1 January 2027. On 22 September OpenAI released GPT-6 Luna at $0.10 and $0.50. On 7 October Anthropic released Claude Haiku 5.5 at the same $0.10 and $0.50 for prompts up to 100,000 tokens, which Anthropic says makes it about 75% cheaper to run than Haiku 4.5 on average, after allowing for its higher token use.
Per token, Gemini 3.8 Flash will cost 15 times as much as either rival from January. Per task the gap is different, because the three models use very different numbers of tokens for the same work. So this article sets independent cost-per-task figures next to list prices, and the worked example treats token usage as a variable.
If your question is what to do about an existing Gemini bill, read our guide to the Gemini 3.8 Flash price increase in January 2027, which covers caching, Batch, Flex and Flash-Lite routing.
Gemini 3.8 Flash: after 1 January
What it is. Google calls gemini-3.8-flash its most intelligent Flash model, built for long-horizon software engineering, autonomous agents and complex enterprise workflows. It is a stable model, sold through the Gemini Developer API and Google Cloud's Agent Platform, and it is the default model for Google's Antigravity managed agent.
Specs. The model page lists a 1,048,576-token input limit and 65,536 output tokens. It accepts text, image, video, audio and PDF input and returns text. Thinking is tunable at low, medium (default) and high.
Pricing from 1 January 2027. $1.50 input and $7.50 output, including thinking tokens. Cached input is $0.15, with cache storage at $1.00 per million tokens per hour and no separate write fee listed. Batch and Flex (preview) are $0.75 and $3.75; Priority is $2.70 and $13.50. Grounding with Google Search gives 5,000 free requests a month across Gemini 3 and newer models, then $14 per 1,000. On Agent Platform, regional endpoints add 10%.
Strengths. It is the only one of the three with native video and audio input, and it ships with Search and Maps grounding, code execution, computer use, file search and URL context. On Vals AI's independent suite it ranks #4 of 138 on GPQA Diamond (94.44%), #4 of 93 on MMMU Pro (89.08%), #3 of 143 on LiveCodeBench (89.48%) and #6 of 35 on Finance Agent v2.1 (65.76%). Artificial Analysis figures reported by OfficeChai on 8 October gave it the highest factual accuracy of the three on AA-Omniscience: 55%, against 44% for Luna and 36% for Haiku 5.5.
Limits. Price, above all. It is also verbose: Artificial Analysis says it generated 130 million tokens running the index and cost $7.77 per index task at the introductory price, a figure that would double at January rates. Its 65,536-token output cap is about half the others'. The same October AA-Omniscience figures showed a 55% hallucination rate, against 40% for Haiku 5.5.
Best for: video and audio understanding, answers grounded in Google Search, and teams whose data, identity and committed spend already sit in Google Cloud.
Claude Haiku 5.5
What it is. Anthropic's smallest current model, claude-haiku-5-5, aimed at high-volume, cost-sensitive work such as summaries, classification, subagents and browser use. It is the first Haiku model with an adjustable effort setting. It is sold on the Claude Platform, Amazon Bedrock, Google Cloud and Microsoft Foundry.
Specs. Anthropic's models overview lists a 1M-token context window, 128,000 max output tokens, adaptive thinking with a default effort of medium, and a reliable knowledge cutoff of June 2026. Input is text and image. Effort runs across five levels, from low to max.
Pricing. For prompts up to 100,000 tokens: $0.10 input and $0.50 output, cache reads $0.01, five-minute cache writes $0.125 and one-hour writes $0.20. Above 100,000 tokens: $0.50 input and $2.50 output, cache reads $0.05, five-minute writes $0.625 and one-hour writes $1.00. Anthropic's pricing documentation counts every input token toward prompt length, cache reads and writes included, and bills a request over the threshold at the higher rates even when part of its prompt is a cache hit. Batch is $0.05 and $0.25 below the threshold, $0.25 and $1.25 above. US-only inference costs 1.1x, on the higher rates too.
Strengths. It leads both independent indexes among these three. Vals AI ranks it #3 of 123 on Vibe Code Bench v1.1 (90.44%) and first of the three on Terminal-Bench 4.0. Anthropic's own launch table, which is vendor-run, reports 72.4% on the OSWorld 2.1 offline subset for computer use, against 48.9% for GPT-6 Luna. Artificial Analysis measures 235 output tokens per second, roughly double the other two.
Limits. Token appetite. Artificial Analysis counts 300 million tokens generated across its index runs, against 110 million for Luna, and Anthropic says the updated tokenizer uses slightly more tokens per task. Vals AI's cost per test is $3.18, seven times Luna's $0.45. The 100,000-token threshold is a cliff: one large retrieved document can push a request to five times the price. OfficeChai also reported on 8 October that Artificial Analysis's figures did not yet reflect the pricing above 100,000 tokens, and that it planned to re-run its AutomationBench test once Anthropic fixed an over-refusal problem.
Best for: coding and research subagents, browser and desktop automation, and agentic knowledge work where prompts stay under 100,000 tokens. Teams pairing it with larger coding models should also read our best LLM for coding in 2027 cost comparison.
GPT-6 Luna
What it is. OpenAI describes gpt-6-luna as its most efficient model for focused, high-volume tasks, released alongside GPT-6 Sol on 22 September 2026. It runs on the Responses, Chat Completions and Batch APIs. The model page lists a 1,050,000-token context window, 128,000 max output tokens and a knowledge cutoff of 18 May 2026. Input is text and image. Reasoning effort runs from none through low, medium (default), high, xhigh and max.
Pricing. $0.10 input, $0.50 output, $0.01 cached input, $0.125 for 30-minute cache writes and $0.25 for 12-hour writes. Prompts above 272,000 input tokens pay 2x input and cache rates and 1.5x output for the whole request, which means $0.20 and $0.75. Batch and Flex are 50% of standard ($0.05 and $0.25), Fast mode is 2x, and regional processing adds 10%.
Strengths. Efficiency. Artificial Analysis puts Luna's cost per index task at $0.15, the lowest of the three, and Vals AI's cost per test is $0.45. Vals ranks it #5 of 47 on CyberBench v1.1 (76.25%). Reasoning set to none gives a cheap non-reasoning mode for classification and extraction, and the Responses API adds hosted web search, file search, code interpreter, computer use and MCP tools.
Limits. It has the lowest index scores of the three (28 on Artificial Analysis v5.0, 52.32% on Vals) and weak hard agentic coding, at 13.64% on Terminal-Bench 4.0 per Vals. The October AA-Omniscience figures showed a 77% hallucination rate. On Chat Completions, function calling works only with reasoning set to none.
Best for: high-volume classification, extraction and routing with short outputs, long-document reads between 100,000 and 272,000 tokens, and budget-capped batch jobs.
Head-to-head comparison
Specs and prices below come from each vendor's model and pricing pages.
| Gemini 3.8 Flash (from 1 Jan 2027) | Claude Haiku 5.5 | GPT-6 Luna | |
|---|---|---|---|
| Released | 2 Sep 2026 | 7 Oct 2026 | 22 Sep 2026 |
| Standard input / output per 1M | $1.50 / $7.50 | $0.10 / $0.50 (prompt up to 100K) | $0.10 / $0.50 (input up to 272K) |
| Long-prompt pricing | Flat to 1,048,576 tokens | $0.50 / $2.50 for the whole request above 100K | $0.20 / $0.75 for the whole request above 272K |
| Cache read per 1M | $0.15 | $0.01 ($0.05 above 100K) | $0.01 |
| Cache write or storage | No write fee listed; storage $1.00 per 1M per hour | $0.125 (5 min), $0.20 (1 hour) | $0.125 (30 min), $0.25 (12 hours) |
| Batch input / output | $0.75 / $3.75 | $0.05 / $0.25 | $0.05 / $0.25 |
| Other tiers | Flex 50% (preview); Priority $2.70 / $13.50 | Batch 50% | Flex 50%; Fast mode 2x |
| Context / max output | 1,048,576 / 65,536 | 1M / 128K | 1,050,000 / 128K |
| Input modalities | Text, image, video, audio, PDF | Text, image | Text, image |
| Reasoning control | low, medium, high | low to max, default medium | none to max, default medium |
| Knowledge cutoff | Not stated on model page | June 2026 | 18 May 2026 |
| Residency option and uplift | Agent Platform regional endpoints, +10% | US-only inference, 1.1x | Regional processing (US, EU), +10% |
| Where to buy | Gemini API, Google Cloud | Claude Platform, AWS, Google Cloud, Microsoft Foundry | OpenAI API |
Published benchmarks, labelled by who ran them
eCorpIT did not run its own tests for this article. The table separates independent benchmarkers from vendors, using the Artificial Analysis and Vals AI model pages and the Anthropic and Google launch posts. Effort settings differ: both independent firms tested Gemini 3.8 Flash at high, its top level, and the other two at max.
| Benchmark | Run by | Gemini 3.8 Flash | Claude Haiku 5.5 | GPT-6 Luna |
|---|---|---|---|---|
| Intelligence Index v5.0 | Artificial Analysis (independent) | 31 | 36 | 28 |
| Cost per index task | Artificial Analysis | $7.77 (intro price) | $0.41 | $0.15 |
| Tokens generated across index | Artificial Analysis | 130M | 300M | 110M |
| Output speed | Artificial Analysis | 120 tokens/s | 235 tokens/s | 118 tokens/s |
| Vals Index | Vals AI (independent) | 55.95% | 58.73% | 52.32% |
| Cost per test (Vals Index) | Vals AI | $5.03 (intro price) | $3.18 | $0.45 |
| Terminal-Bench 4.0 | Vals AI | 19.70% | 36.36% | 13.64% |
| Finance Agent v2.1 | Vals AI | 65.76% | 65.37% | 56.84% |
| Terminal-Bench 4.0 | Anthropic (vendor) | Not reported | 39.2% | 16.4% |
| OSWorld 2.1, offline subset | Anthropic (vendor) | Not reported | 72.4% | 48.9% |
| HLE-Verified | Google (vendor) | 54.9% | Not reported | Not reported |
Three cautions. Artificial Analysis launched Index v5.0 on 27 October 2026, so its scores are not comparable with earlier coverage, such as the 59 it gave Gemini 3.8 Flash at launch on the older scale. OfficeChai's 8 October report said Artificial Analysis cost figures then excluded Haiku 5.5's pricing above 100,000 tokens. And the Anthropic rows use Anthropic's own harness and omit Gemini; OpenAI-run figures are not included.
Worked cost example
Two illustrative scenarios using only vendor list prices. Measure your own prompt sizes with our free LLM token counter, and remember the three tokenizers produce different counts for the same text.
Scenario A: a RAG assistant at 1 million requests a month
Each request sends 8,000 input tokens: a 5,000-token stable prefix and 3,000 dynamic tokens. Each returns 700 output tokens. With caching, 80% of the prefix (4,000 tokens) is a cache hit; the missed 1,000 prefix tokens are billed as cache writes on Anthropic and OpenAI and as plain input on Gemini.
| Monthly cost | Gemini 3.8 Flash (Jan 2027) | Claude Haiku 5.5 | GPT-6 Luna |
|---|---|---|---|
| No caching: 8,000M input + 700M output | $12,000 + $5,250 = $17,250 | $800 + $350 = $1,150 | $800 + $350 = $1,150 |
| With caching: 4,000M cached | $600 | $40 | $40 |
| Missed prefix: 1,000M | $1,500 (input rate) | $125 (write rate) | $125 (write rate) |
| Dynamic input: 3,000M | $4,500 | $300 | $300 |
| Output: 700M | $5,250 | $350 | $350 |
| Total with caching | $11,850 | $815 | $815 |
| Batch, no caching | $8,625 | $575 | $575 |
On list prices, Haiku 5.5 and Luna tie; token usage breaks the tie. If Haiku 5.5 needed 2.7 times Luna's output tokens on your task (the 300 million to 110 million ratio Artificial Analysis recorded across its index), its output line would rise from $350 to $945 and its total from $815 to $1,410. Short-answer workloads at low effort should sit well below that ratio, which is why you test. On Gemini 3.8 Flash the cached workload costs $5,925 today and $11,850 from January.
Scenario B: long-document review at 3,000 requests a month
Each request sends one large contract pack with no caching and returns 3,000 output tokens.
| Prompt size | Gemini 3.8 Flash | Claude Haiku 5.5 | GPT-6 Luna |
|---|---|---|---|
| 150,000 tokens (450M input, 9M output) | $675 + $67.50 = $742.50 | $225 + $22.50 = $247.50 | $45 + $4.50 = $49.50 |
| 300,000 tokens (900M input, 9M output) | $1,350 + $67.50 = $1,417.50 | $450 + $22.50 = $472.50 | $180 + $6.75 = $186.75 |
At 150,000 tokens Haiku 5.5 is past its threshold and Luna is not, so Luna costs a fifth as much. At 300,000 tokens both surcharges apply and Luna still costs under half. Gemini stays flat per token but remains the most expensive.
Decision table: which cheap model to pick
| Pick | If |
|---|---|
| Claude Haiku 5.5 | You run subagents, browser automation, coding subtasks or knowledge work, prompts stay under 100,000 tokens, and you want the top independent index score in this tier |
| GPT-6 Luna | Cost per task matters most, outputs are short (classification, extraction, routing), prompts fall between 100,000 and 272,000 tokens, or you use OpenAI's Responses API tools |
| Gemini 3.8 Flash | Inputs include video or audio, you need Google Search or Maps grounding, your data and identity controls sit in Google Cloud, or your evals show it wins on your documents |
| Gemini 3.5 Flash-Lite | You must stay on Google but want a lower bill: $0.30 input and $2.50 output, with no dated change |
| More than one | Volumes are large and workloads differ: route per task through a gateway and keep a second vendor as fallback |
Gemini 3.5 Flash-Lite pricing is from Google's pricing page. A common design pairs a larger planner with Haiku 5.5 subagents, or puts a cheap Luna classifier in front to pick the model per request. Our model routing framework explains how to classify workloads first.
Data residency: India, EU and US
- Gemini 3.8 Flash. Agent Platform regional endpoints cost 10% more, $1.65 and $8.25 from January. We found no Google page confirming a Mumbai (asia-south1) API endpoint for Gemini 3.8 Flash as of 1 November 2026. Google's data residency page for the Gemini Enterprise app says its India in-country region does not offer 3.8 Flash and routes such traffic to the global endpoint after an administrator accepts a warning. A 21 September 2026 post on Google's developer forum, from an Indian team evaluating Gemini Flash-Lite models for a voice agent, reported only global and us-central1 options.
- Claude Haiku 5.5. Anthropic offers US-only inference at 1.1x, including on the above-100,000-token rates. Partner-operated platforms such as Amazon Bedrock set their own regional pricing, so check Haiku 5.5's regions there for other locations.
- GPT-6 Luna. EU data residency covers Standard, Fast, Flex and Batch. OpenAI's data residency guide lists India for storage at rest, but regional processing only in the US and Europe, and residency outside the US requires approval for its abuse monitoring controls plus a modified retention amendment. Regional processing carries a 10% uplift.
Under India's Digital Personal Data Protection Act 2023, the data fiduciary stays responsible for processing done on its behalf and may use a data processor only under a valid contract. Pin personal-data workloads to vendors and regions you have assessed, and let any cost-based router choose only among those.
Migration checklist: moving traffic between these models
- Build an eval set of 200 to 500 real requests with expected outputs or grading rubrics.
- Run each candidate at two effort levels; record quality, token counts and latency, since tokenizers differ.
- Calculate cost per successful task, not cost per token.
- Check your prompt-size distribution against Haiku's 100,000-token and Luna's 272,000-token thresholds.
- Rework caching. Anthropic's five-minute and OpenAI's 30-minute cache writes cost 1.25x input, and their longer-lived writes cost 2x and 2.5x, while Gemini's implicit caching has no write fee but needs requests of at least 4,096 tokens.
- Put providers behind a gateway so a model change is configuration, not code; our AI gateway comparison covers the options.
- Shadow production traffic, canary a small share with fallback, and review data-processing terms before personal data moves.
How eCorpIT can help
eCorpIT is a Gurugram-based technology organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with senior-led engineering teams working across AWS, Microsoft and Google platforms. Our AI gateway and model routing service builds eval sets from your real traffic, measures cost per successful task across Gemini, Claude and OpenAI models, and sets up routing with fallbacks so a model change does not need a code release. Talk to us at /contact-us/.
Last updated: 1 November 2026.
