Alan West

Posted on Nov 9 Originally published at blog.authon.dev

Gemini Omni 1.2 Flash vs GPT-6.1 Terra vs Claude Sonnet 5.5: Omni Is Video-Only

#ai #programming #productivity #webdev

If you came here to find out whether Gemini Omni 1.2 Flash is cheaper than GPT-6.1 Terra or Claude Sonnet 5.5 for your API calls, you're comparing the wrong models. Omni 1.2 Flash is a video generator. It makes 3 to 15 second clips at $0.07 per second of 720p. It won't write your SQL migration.

The comparison you probably want is Gemini 3.8 Flash, Google's current Flash text model, against Terra and Sonnet 5.5. Short version: 3.8 Flash is the cheapest per agent task and the most expensive per solved task, its price doubles on January 1, 2027, and Terra comes out ahead on cost per solved task. Numbers below.

I opened Google's launch post expecting a Flash-tier text model. What I found was a video model with "Flash" in the name, and an HN thread where one of the top complaints is exactly that. User ricardobeat: "The 'omni' naming for a video-only model is so confusing."

So I'll do both: the real numbers for Omni 1.2 Flash first, then Gemini 3.8 Flash vs GPT-6.1 Terra vs Claude Sonnet 5.5 on price, Terminal-Bench 4.0 and cost per task. Every figure comes from official pricing pages, launch posts and public leaderboards. I didn't run any of the benchmarks myself.

Gemini Omni 1.2 Flash pricing and rate limits

From the Gemini API pricing page:

	Omni 1.2 Flash	Omni 1.1 Flash
Input (text/image/video), per 1M tokens	$1.50	$1.50
Video output, per 1M tokens	$12.00	$17.50
720p, per second	~$0.07	~$0.10
10s clip at 720p	~$0.70	~$1.00
4K (upsampled), per second	~$0.209	n/a
Free tier	none	none

Google's "30% cheaper" claim holds up on the API: $17.50 down to $12.00 per million video tokens is a 31% cut. Two things the launch post leaves out:

Flow got more expensive. HN user cube00 reports a 720p 10s clip went from 15 credits on Omni 1.1 to 30 on 1.2. Another commenter says the 50 free daily credits now buy about one video a day.
The rate limits are tiny. The rate limits page lists gemini-omni-1.2-flash at 8 RPM and 40 RPD on Tier 1, with the free tier "Not supported." Forty clips a day is fine for prototyping and nowhere near the "cost-effective at scale" pitch in the announcement.

Also, starting December 3, 2026, requests to Omni 1.1 Flash get routed to 1.2 and billed at 1.2 rates. If you pinned 1.1 for consistency, your outputs change next month whether you touch anything or not.

Early quality reports are mixed. One commenter calls the character consistency gain "neglegible" and got a bike "moving in reverse." Another says it's "still pretty inconsistent, slow, and expensive." The first one says they're having a blast with Wan 3.0 instead, and a third calls Flow "a joke" next to Seedance or MiniMax.

Gemini 3.8 Flash vs GPT-6.1 Terra vs Claude Sonnet 5.5: API prices

	Gemini 3.8 Flash	GPT-6.1 Terra	Claude Sonnet 5.5
Input / 1M	$0.75 (→ $1.50 Jan 1, 2027)	$1.00	$2.00
Cached input / 1M	$0.075 (→ $0.15) + storage	$0.05	$0.10
Output / 1M	$3.75 (→ $7.50 Jan 1, 2027)	$5.00	$10.00
Batch	50% off ($0.375 / $1.875)	50% off ($0.50 / $2.50)	50% off ($1 / $5)
Context	1M in, 64K out	~1.05M, 128K out	1M at standard price
Long-context surcharge	none listed	above 272K input	none
Free tier	yes	no	no

Sources: Gemini pricing, Gemini 3.8 Flash model card, Claude pricing, and for Terra, OpenAI's launch as reported by Interesting Engineering and APIMaster's spec table.

The Gemini column needs a footnote. $0.75/$3.75 is the rate through December 31, 2026, and it doubles on January 1, 2027. After that, 3.8 Flash costs $1.50/$7.50, more than Terra on both input and output. If you switch to it for price, you're signing up for a 2x bill in under two months.

Terminal-Bench 4.0: Gemini 3.8 Flash scores 19%, not 89%

Every vendor picks its own benchmarks, but one shows up for all three: Terminal-Bench 4.0. Here's what each vendor reports for its own model, next to the independent Vals AI run mirrored on BenchLM (snapshot November 5, 2026):

Model	Vendor-reported	Vals AI
Claude Sonnet 5.5	70.6%	64.14%
GPT-6.1 Terra	50.3% (max effort)	49.49%
Gemini 3.8 Flash	19.1%	19.70%
Claude Haiku 5.5	n/a	36.36%

Google's model card leads with 89.4% on Terminal-Bench 2.1, and that's the number most coverage repeats. The same card lists 19.1% on Terminal-Bench 4.0. Those are different versions of the benchmark and the new one is much harder. On 4.0, 3.8 Flash lands below Haiku 5.5, which costs $0.10/$0.50 for prompts up to 100K tokens.

Credit to OpenAI: Terra's self-reported number is within a point of the independent one. Anthropic's 70.6% is about six points more generous than Vals, and Artificial Analysis measured 63.6% (on a pre-release build, per Decrypt). Sonnet is still on top either way.

On desktop control (OSWorld 2.1), OpenAI's own numbers put Sonnet 5.5 at 83.9% versus Terra's 74.2%. The same numbers have Terra at $1.19 per task against Sonnet's $5.74. Nobody has reproduced those independently yet.

Cost per solved task: where Flash loses

My own arithmetic for one coding-agent task: 500K input tokens, 450K of them cache hits, plus 30K output tokens. That's roughly a multi-turn session in a medium-sized repo. I applied each model's list price to the same token counts. Real models spend different numbers of tokens on the same task, so treat these as ballpark.

Model	Cost per task	Vals TB 4.0	Cost per solved task
Gemini 3.8 Flash (until Dec 31)	$0.18	19.7%	~$0.93
Gemini 3.8 Flash (from Jan 1)	$0.37	19.7%	~$1.87
GPT-6.1 Terra	$0.22	49.5%	~$0.45
Claude Sonnet 5.5	$0.45	64.1%	~$0.69
Claude Haiku 5.5	$0.02	36.4%	~$0.07

Math for Terra, so you can check me: 50K × $1 + 450K × $0.05 + 30K × $5 = $0.05 + $0.0225 + $0.15 ≈ $0.22. These figures leave out cache-write premiums and Gemini's cache storage fees, and both push real costs up.

The Haiku row assumes every request stays at or under 100K prompt tokens. Above that, Haiku 5.5 bills at $0.50/$2.50 with $0.05 cache hits, which puts this task at about $0.12, or roughly $0.34 per solved task. That's still the cheapest row, just by a smaller margin.

Per task, Flash looks like the bargain. Divide by how often it actually finishes the job and it's the most expensive of the big three, before and after January 1, because you pay for the failed runs too.

Which model should you use for cheap API calls?

My take, by workload:

For agent loops and terminal work, pick GPT-6.1 Terra if cost per solved task is what you care about, and Sonnet 5.5 if failures are expensive (human review time, broken CI). Sonnet roughly doubles the per-task bill for about 15 points of success rate.
For high-volume classification, extraction and rewriting, skip all three. Haiku 5.5 at $0.10/$0.50 or Gemini 3.5 Flash-Lite at $0.30/$2.50 is the tier you want.
For prototyping on zero budget, Gemini 3.8 Flash's free tier is real and useful. Just don't build a cost model on the December price.
For video, Omni 1.2 Flash is a fair deal on the API at $0.70 per 10s clip, if 40 requests a day is enough.

Before you move anything, check the bits the launch posts skip:

# Which Gemini model IDs are you actually calling?
grep -rn "gemini-omni-1.1-flash\|gemini-3.8-flash" src/

# Put a reminder in for the Jan 1, 2027 Gemini price change
echo "Gemini 3.8 Flash doubles to \$1.50/\$7.50 on 2027-01-01" >> COSTS.md

If you route through OpenRouter, one HN commenter said their calls kept failing there with Gemini uptime at 68%. Put retries and timeouts into your cost estimate before you trust the per-token price.

Which model are you running for cheap, high-volume calls right now, and would these numbers make you switch? And is anyone getting better than 20% out of Gemini 3.8 Flash on real agent work?

DEV Community