Three frontier APIs shipped or changed pricing within a week of one another, landing in three distinct price bands. Google released Gemini 3.8 Flash on September 2, 2026, at $0.75 per million input tokens and $3.75 per million output tokens. Anthropic released Claude Fable 5.1 the day before at $10 and $50. OpenAI’s GPT-5.6 Sol sits between them at $5 and $30. On Artificial Analysis’ independent Intelligence Index, they score 59, 57, and 59 respectively. A model priced at roughly one-thirteenth of the top tier matches it on the only index that evaluates all three the same way.
The catch is the word “index.” Benchmarks count answers per task; your bill counts tokens per task. Gemini 3.8 Flash is designed to spend more of them.
This guide compares the three models by:
Google’s launch post is the primary source for the Gemini claims below. For context, our August comparison used Claude Fable 5, not 5.1. Anthropic replaced that model on September 1, making this a new comparison rather than a refresh.
| Specification | Gemini 3.8 Flash | Claude Fable 5.1 | GPT-5.6 Sol |
|---|---|---|---|
| Vendor | Anthropic | OpenAI | |
| Context window | 1,048,576 tokens | 1M tokens | 1M tokens |
| Maximum output | 65,536 tokens | 128K tokens | 128K tokens |
| Input price per 1M tokens | $0.75 introductory; $1.50 from Jan. 1, 2027 | $10 | $5 |
| Output price per 1M tokens | $3.75 introductory; $7.50 from Jan. 1, 2027 | $50 | $30 |
| Cache read per 1M tokens | $0.075 introductory; $0.15 from Jan. 1 | $0.25 | $0.50 |
| Knowledge cutoff | March 2026 | June 2026 | Not listed here |
| Reasoning control | |||
thinking_level : low, medium, high; medium default |
|||
| Extended thinking, always on | Effort levels up to xhigh |
||
| Artificial Analysis Index | 59 at high | 57 at medium | 59 at xhigh |
Two specifications matter immediately:
The Gemini 3.8 Flash pricing guide covers the doubling, caching, batch, and grounding rates in detail.
Artificial Analysis runs the same nine evaluations against each model and publishes one Intelligence Index score:
xhigh
effort: Gemini 3.8 Flash improved from 56 for 3.7 Flash and 52 for 3.6 Flash. Fable 5.1’s score of 57 also matches GPT-5.6 Terra at max effort and Muse Spark 1.2 at xhigh
. Grok 4.6 at medium is another model at 59.
Read the reasoning settings before interpreting those numbers. Fable 5.1 was tested at medium, while Sol was tested at its top listed setting, xhigh
. A two-point difference across different effort levels is not a definitive ranking.
The defensible conclusion is narrower: at the tested settings, Gemini 3.8 Flash matches the two premium models on this index and is the cheapest model with a score of 59 by a wide margin.
Artificial Analysis also measured token usage and latency for Gemini 3.8 Flash at high thinking:
The model thinks before it writes, which explains the relatively high time to first token. Keep these measurements in mind when the per-token price appears too attractive.
Google’s Flash page publishes three cross-vendor benchmark rows in text form, but Claude Fable 5.1 is not included. The Anthropic entries are Claude Opus 5 and Claude Sonnet 5 instead. Fable 5.1 had been public for only one day when Google’s tables were published.
The Anthropic results below are therefore proxies, not direct results for Fable 5.1:
| Google-run benchmark | Gemini 3.8 Flash | Claude Opus 5 | Claude Sonnet 5 | GPT-5.6 Sol | GPT-5.6 Terra |
|---|---|---|---|---|---|
| HLE-Verified | 54.9% | 54.4% | 31.0% | 54.5% | 51.1% |
| Vals Finance Agent v2 | 61.4% | 58.6% | 53.9% | 53.8% | 54.4% |
| Harvey Legal Agent Benchmark | 10.0% | 6.7% | 5.0% | 2.5% | 0.8% |
HLE-Verified is effectively a three-way tie: Gemini 3.8 Flash, Opus 5, and Sol are within half a point of one another.
Vals Finance Agent v2 is where Gemini 3.8 Flash separates:
Harvey Legal is the outlier. Every model scores below 11%, so the ordering is more meaningful than the gaps.
Google did not publish text results for:
The DeepSWE v1.1 result appears only as an image-table claim that Gemini 3.8 Flash “outperforms most larger frontier models” at “a fraction of the cost.”
For coding and computer-use comparisons, you must rely on separate vendor-run evaluations, which do not overlap. Anthropic’s results are in the Claude Fable 5.1 benchmarks breakdown, and OpenAI’s are in the GPT-5.6 Sol benchmarks. Artificial Analysis remains the only like-for-like source in this comparison.
At introductory rates:
Cache reads narrow the difference:
Fable 5.1’s cache-read price is half of Sol’s—the one pricing line where the most expensive model is not the most expensive option.
Assume a request uses 1 million input tokens and 200,000 output tokens, with no cached input:
$0.75 + $0.75 = $1.50
$5.00 + $6.00 = $11.00
$10.00 + $10.00 = $20.00
From January 1, 2027, the same Gemini request costs $3.00. The gaps then compress to:
The price doubling also applies to Gemini 3.6 Flash and 3.7 Flash, so there is no cheaper Gemini Flash fallback at the same introductory rates.
See the Claude Fable 5.1 pricing guide and GPT-5.6 pricing guide for batch and cache tiers.
Google says Gemini 3.8 Flash “can use more tokens on longer running and complex tasks, by design.”
On difficult problems, the model may:
Artificial Analysis measured the impact:
| Gemini 3.8 Flash setting | Cost per task |
|---|---|
| High | $0.58 |
| Medium | $0.41 |
| Low | $0.24 |
At high, Gemini used an average of 48,000 output tokens per task—30% more than Gemini 3.7 Flash. As a result, its measured cost rose from $0.40 per task for 3.7 Flash to $0.58 for 3.8 Flash at unchanged token prices.
Two conclusions follow:
thinking_level
is a major cost control. Moving from high to low reduced measured cost by more than half.Artificial Analysis has not published comparable per-task figures for Fable 5.1 or Sol in the available text. Benchmark your own prompts before assuming any multiplier.
The Gemini 3.8 Flash thinking-levels guide explains how to configure the setting per endpoint. In some routes, Gemini 3.7 Flash may still be the better buy because its measured per-task cost is 31% lower; see the 3.8 Flash vs. 3.7 Flash comparison.
Use Gemini 3.8 Flash as the default for thousands of agent tasks per day. It matches the premium models on the independent index, leads Google’s finance and legal agent rows, and remains substantially cheaper per token after the January 2027 price increase.
Additional advantages include:
Trade-offs:
Set thinking_level
to medium or low where the workload does not require high-depth reasoning.
Fable 5.1 is an escalation model rather than the default starting point. Use it when an incorrect result costs more than the additional tokens, such as:
Its 128K output limit and $0.25 cache reads are useful for prefix-heavy agent loops that reuse a large context on every turn. In those workloads, the effective cost multiplier can be much smaller than 13.3×.
See what Claude Fable 5.1 is for its own specification and positioning.
Sol scores 59 at xhigh
, ties Gemini 3.8 Flash on HLE-Verified, and offers 128K output at $5 input and $30 output per million tokens.
It is a practical premium option when your agents, evaluations, and tooling already use OpenAI’s ecosystem. The downside is its $0.50 cache-read price, the highest of the three. Sol is therefore a better fit for fresh-context workloads than for long cached sessions.
The Grok 4.6 vs. GPT-5.6 vs. Claude Fable 5 comparison provides additional context on Sol versus the previous Anthropic flagship.
Every model argument should end with the same action: measure your own prompts.
Apidog is an API client and testing platform. It does not run these models, but it makes it fast to send the same request to all three providers and compare the results.
Create one project with three environments. Store each provider’s base URL and API key as environment variables:
| Environment | Base URL variable | API key variable |
|---|---|---|
| Gemini | https://generativelanguage.googleapis.com |
|
GEMINI_API_KEY |
||
| Claude | Anthropic’s base URL | Your Claude key |
| OpenAI | OpenAI’s base URL | Your OpenAI key |
Keep keys in environment variables—not in request bodies or shared collections.
Add three request steps with the same prompt.
The Gemini request uses /v1beta/interactions
with:
{
"model": "gemini-3.8-flash",
"thinking_level": "medium"
}
The Claude and OpenAI steps should use their respective messages or chat endpoints.
Add assertions to every step:
200
For Gemini’s legacy endpoint, the reasoning-token field is:
usageMetadata.thoughtsTokenCount
Assert against the equivalent usage block for the other providers.
Run the scenario against a set of golden prompts, then schedule it daily. If a vendor changes a default or a model starts consuming more thinking tokens, the failed assertion alerts you before the invoice does.
Use the AI agent API testing guide for multi-turn scenarios and scheduling API tests in Apidog for recurring runs.
Download Apidog to create the three at $0.75 input and $3.75 output through December 31, 2026.
Artificial Analysis measured its cost per task at:
Because token usage varies by model and prompt, measure cost per completed task—not just cost per token.
Yes:
Maximum output differs. Gemini supports 65,536 tokens, while the other two support 128K.
Google’s published rows list Claude Opus 5 and Claude Sonnet 5. Fable 5.1 shipped on September 1, one day before Gemini 3.8 Flash, so it was not included.
For Anthropic’s own results, see the Claude Fable 5.1 vs. Opus 5 comparison.
Partly. On January 1, 2027, Gemini 3.8 Flash moves to $1.50 input and $7.50 output per million tokens.
It remains cheaper, but the gap falls to:
Start with Gemini 3.8 Flash for high-volume workloads. Set thinking_level
per route and monitor tokens per task.
Escalate to Claude Fable 5.1 when cheaper models fail your evaluations, especially when you reuse a large context across turns. Choose GPT-5.6 Sol when the rest of your stack is already OpenAI-based and you want a premium model without adding another vendor.
Before committing, send the same golden prompts to all three models in one test scenario. The price ratio is public; the cost-per-task ratio for your prompts is yours to measure.