Back to Blog

Gemini 4 Neon vs. GPT-6.2 Astra vs. Claude Fable 5.2 for AI Agents

November 21, 2026

Google charges $15 per million output tokens for Gemini 4 Neon. OpenAI charges $40 for GPT-6.2 Astra. Yet on the Artificial Analysis Intelligence Index, Astra costs $4.00 per task to Neon's $6.44. Give the same models the professional assignments in the Vals Index and the order flips: about $7.40 for Neon, $10.68 for Astra. If you're choosing a model for an agent, its job description tells you more about the bill than the token price does. Google launch notes · OpenAI launch notes · Artificial Analysis · Vals results for Neon and Astra

Google released Neon on November 18, 2026, positioning it between Hydrogen for everyday execution and Carbon for long projects. Its model card compares Neon mainly with GPT-6.2 Sol and Claude Opus 5.5. We'll put it up against two flagships a team can actually deploy, GPT-6.2 Astra and Claude Fable 5.2, and keep Sol as a budget reference. Anthropic's other new model, Mythos 5.2, has additional access restrictions. DeepMind model card · OpenAI model lineup · Anthropic launch notes

What the three models cost to put to work

As of November 21Gemini 4 NeonGPT-6.2 AstraClaude Fable 5.2
Standard API per million input / output tokens$3 / $15$8 / $40$8 / $40
Per million cached input tokens, read hits$0.15$0.40$0.20
Artificial Analysis Intelligence Index616265
Same index, cost / elapsed time per task$6.44 / 22.3 min$4.00 / 14.2 min$26.54 / 42.2 min
Vals Index score / cost per test70.87% / about $7.4067.66% / $10.6871.31% / $17.94

The API figures are base rates in US dollars. Cache writes, long requests and fast modes have their own terms. Artificial Analysis used xhigh for Neon and max for Astra and Fable, with default fallback for Fable. Compare costs within a row: the two labs use different tasks and accounting. Fable's 0.44-point lead over Neon on Vals is smaller than either model's roughly 0.95-point reported margin. That's a slim basis for paying a premium across every assignment. Google rates · Astra billing · Fable pricing · AA leaderboard · Vals: Neon, Astra and Fable

Take a single call with 100,000 uncached input tokens and 10,000 billable output tokens. At base rates that costs $0.45 on Neon, $1.20 on either flagship, and $0.30 on Sol. The arithmetic assumes identical usage. An agent might reread the repository twice, run another test or need a second attempt. For a production budget, divide total spend across all attempts by the number of tasks that pass acceptance. That's your cost per successful task. Sol pricing · Holistic Agent Leaderboard research · Budget-constrained agent evaluation

Coding agents: a passing terminal session or a patch you'd merge?

On the public Terminal-Bench 4.1 leaderboard, Fable 5.2 in Claude Code resolves 65.8% ± 2.8 of tasks, Astra in Codex resolves 62.1% ± 2.3, and Neon in Antigravity CLI resolves 61.2% ± 2.8. The margins are 95% confidence intervals in percentage points. The settings are max, max and xhigh, respectively. These results travel with the CLI, its tools and its context management. If you've seen 65.2% for Neon in Google's model card, its methodology traces that figure to a separate Artificial Analysis run with Antigravity CLI 1.5.9. The three figures above all come from Terminal-Bench's own board. Public leaderboard · Version 4.1 notes · Google evaluation methodology

Cognition's FrontierCode gives a better feel for what a maintainer buys. One JavaScript task involves resetting a generator in a mocking library. A patch can fix the visible symptom while quietly skipping cleanup code inside the user's finally block. That's the kind of ticket I'd use to compare Fable and Astra on an existing codebase. Time the review and count how often the engineer has to send the patch back. A team already working in Antigravity has a different reason to try Neon: it can plan the change, delegate scoped edits and tests to Hydrogen, and take over again when a worker gets stuck. FrontierCode task and grading · Antigravity's planner and executor

An earnings filing, a chart and a recording in one request

A research agent might receive a footnoted earnings filing, an investor slide and a recording in which management explains a change in segment reporting. I'd start that job on Neon. It accepts text, images, audio and video natively. Its 1,048,576-token window is shared by inputs, reasoning, tool calls and output, so reserve space for the answer before loading the last attachment. Astra's API lists a 922,000-token input limit and a 128,000-token maximum output. Its standard rates also change above 272,000 input tokens. Neon model card · Google developer guide · Astra API limits and rates

Vals gives that choice some empirical backing. Neon scored 68.32% on its Finance Index and 78.36% on its Legal Index. Astra took first place on its Science Index. A useful acceptance question for the earnings task is: On which page did the company restate last year's segment figures, and which year-over-year comparison changes as a result? Ask for the page, the passage and the recalculation. If the daily assignment instead involves running an analysis program, testing a hypothesis and interpreting the output, I'd put Astra first on the shortlist. Neon results · Finance Index design · Astra results · Science Index design

For an agent that takes action, how much thinking should you buy?

Zapier's AutomationBench checks the final state of a simulated business. Imagine an agent asked to update an order and notify a customer. A plausible confirmation message is easy to produce. The order and the notification both have to end up in the right state. Across 657 tasks, Neon's medium setting passed 53.3% at an estimated $0.58 per task. High reached 54.3% at $0.88. Xhigh reached 56.3% at $1.69. Moving from high to xhigh bought two percentage points for an extra $0.81 per attempt. For a busy operations queue, test medium or high first, then identify the exceptions that justify escalation. Zapier release report · AutomationBench scoring

One more test is worth writing. Halfway through a run, tell the agent, “Switch to the test account and hold the email.” Check which account it uses and whether a message actually goes out. Arena calls the ability to absorb a mid-task correction agent steerability. For a personal assistant with access to an inbox, this can matter more than shaving a minute off a completed task. Arena signal definitions · Arena evaluation method

Which API would we connect first?

Neon would get the first batch of mixed-media research and cross-app operations jobs. Astra would get the first scientific analysis jobs and a trial on engineering work already running in Codex. Fable earns a place in the first round when a maintainer has to approve every patch and rework is expensive. For simpler, high-volume calls, keep Sol in the comparison: it costs $2 per million input tokens and $10 per million output tokens, and recorded the lowest cost per task of these four on the Artificial Analysis table. Vals · Astra · Terminal-Bench · Sol · Artificial Analysis

Before committing, take 20 jobs you've already delivered and can grade. Give each candidate the same tools, permissions and retry allowance. Record first-pass acceptance, total spend per accepted job and minutes of human revision. For an agent whose work somebody has to sign, that last column can settle the decision.

Based on public evaluations, model documentation and standard API rates available November 21, 2026. Deployment choices are editorial judgments.

Related Articles