Gemini 4 Neon vs Claude Opus 5.5: 6 Coding Tests, Time and Cost
Prefer to watch? Watch on YouTube · Subscribe for new builds · Read the transcript
In this video I put Google's new Gemini 4 Neon up against Claude Opus 5.5 on six tests: three AAA style browser games, a client app with hidden tests, a website design job, and our hardest bug hunt. I show the time and cost for every run. It's for anyone who codes with Opus every day and wants to know if Neon is worth switching to.
What I cover in this video
- 0:45 What Gemini 4 Neon is: where it sits in the Gemini 4 family, the price, the context window, and Google's own launch table.
- 3:03 Why the high setting matters: Neon at high is close to its max, and Opus at high is well under its max.
- 7:00 Why I ran Neon inside Claude Code through OpenRouter, and the 15 minute restart rule.
- 8:32 Game 1, the city flight: Neon's neon city next to the black screen from Opus.
- 10:04 Game 2, the railgun robot, where the Opus build was the better game to play.
- 12:22 The client app: an accounting firm intake page graded by 12 hidden tests.
- 14:40 The website design job: hard checks, two vision judges, and a side by side.
- 18:34 The verdict: all six results together and who should switch.
What Gemini 4 Neon is and what the benchmarks say
Google DeepMind released Gemini 4 Neon on November 18. Google says it's adapted from Gemini 4 Carbon and sits in the middle of the family. Carbon is for the hardest, longest projects, Hydrogen is the fast everyday model, and Neon balances depth and efficiency. It has a 1 million token context window and four thinking levels: low, medium, high, and extra high. High is the default.
The price is $3 in and $15 out per million tokens. Opus 5.5 is $4 in and $20 out, so Neon is 25% cheaper per token.
Google's launch table compares Neon directly to Opus 5.5. Neon is higher on 16 rows, Opus on 2, and one row has no Opus score. The two rows Opus wins are Terminal-Bench 4.1 and FrontierCode 2.0, the agentic coding rows, and it wins them by about a point. Those are Google's own numbers, so I also looked at the independent ones:
- Artificial Analysis Intelligence Index: at top settings, Neon scores 61 and Opus 59, at $6.44 per task against $10.04. At the high setting, Neon scores 60 and Opus 52.
- Vals Index: Neon 70.87, Opus 70.14. That's inside their error bars, so it's a tie, but Neon cost $7.40 per test against $24.77 for Opus.
- Arena: as of November 20, Opus at its max setting leads the agent leaderboard 1336 to 1333, and leads the webdev board 1819 to 1765.
How I set up the tests
Both models ran as coding agents inside Claude Code. Same prompt, a fresh folder, high effort, web search off, one shot, and a 45 minute limit (60 for the bug hunt). Opus ran on my Claude subscription, so its cost is the API equivalent price. Neon ran through OpenRouter, so its cost is what we were billed.
I didn't use Google's own agent because the Antigravity command line on our Google account offers no Neon. Claude Code is Anthropic's home field, and Google's agent might get more out of Neon. Opus normally runs on medium, and I bumped it to high so both had the same setting name.
One rule applied to both: if a model hasn't saved anything 15 minutes into a build, we restart it once in the same folder and the clock keeps running. Scored tests are graded by hidden tests or scripted checks the models never see, plus two judge models from labs that aren't in this matchup. Each result is one finished run, so treat it as a vibe check and not a final verdict.
The three browser games
The prompts asked for AAA production values: real lighting and shadows, bloom, particles, physics, a HUD, and sound. Everything had to be generated in code, with no downloaded models or textures, in one HTML file that runs in the browser. Every game was played with the exact same keyboard inputs.
- City flight: Neon finished in 19.7 minutes for $2.70 and it plays, with hundreds of towers, rain, glowing rings, and a crash and respawn. Opus was restarted at 15 minutes and hit the 45 minute limit, and its saved version shows a black screen. An unscored Opus retry finished in 28.3 minutes for $6.53 and plays.
- Railgun robot: Neon took 17.4 minutes and $2.39. Opus took 37.8 minutes and $4.84 with a restart. On the same 30 seconds of inputs, the Opus robot destroyed 6 targets and Neon's destroyed 1.
- Coastal rally: Neon finished in 13.9 minutes for $2.12. Opus was restarted and hit the 45 minute limit, but its saved version plays.
The game test goes to Neon. It built three playable games out of three, every one under 20 minutes, for about $7 total. Opus had to be restarted on all three, though when it did finish, its robot game was the better one to play.
The real work tests and who should switch
Client app. A two partner accounting firm wants a page where clients upload tax documents, and the request comes as a messy email thread with a requirement that changes partway through. Both models passed 12 of 12 hidden tests, and both reply emails scored 10 of 10 from both judges. Neon took 10.7 minutes and $1.56. Opus took 4.6 minutes and $1.24.
Website design. A one page site for a bike repair shop. Neon passed 5 of 5 hard checks. Opus passed 4 of 5 because some of its orange text fell just under the contrast line. Both vision judges picked Opus, 111 to 109 out of 120. Neon took 11.5 minutes and $1.47, Opus 8.1 minutes and $2.29. Both models picked almost the same palette and both built the design around a repair tag.
Bug hunt. An invoicing backend with 16 hidden bugs and 11 decoys. Both scored 28 of 32, fixed 14 of 16, broke nothing, reported no fake bugs, and missed the same two bugs. Opus took 8.2 minutes and $2.51. Neon took 17.4 minutes and $2.92.
Google's table said 16 to 2. On our tests the quality was basically even, and the real difference was speed. On the big game builds, Opus sat thinking past the 15 minute mark on all three scored runs. On the client job and the bug hunt, Opus finished in less than half the time.
Key takeaways
- If you build visual stuff like games, 3D scenes, demos, and prototypes, Neon is the real deal and it's cheap per token. All six tests billed $13.15.
- If most of your day is client work and fixing bugs, I wouldn't switch just for this. The quality was the same and Opus got there faster.
- If you already pay for Google AI Pro or Ultra and work in Antigravity, Neon is worth a real try. Watch your usage limits.
- We ran both on high. For Neon that's basically its max, and Artificial Analysis has Opus at max 7 points above its high score.
- Opus is still my daily driver. I'm on the $200 plan, so I'm not paying per token, and the work that pays my bills looks more like the client job than a neon city.
FAQ
Is Gemini 4 Neon better than Claude Opus 5.5 for coding?
In my six tests the quality was basically even. Neon won the three game builds, the client app and bug hunt were ties, and the design job was split. The real difference was speed, and it flipped depending on the job.
How much does Gemini 4 Neon cost compared to Opus 5.5?
Neon is $3 in and $15 out per million tokens, and Opus 5.5 is $4 in and $20 out, so Neon is 25% cheaper per token. Neon's bill for all six of my tests came to $13.15 through OpenRouter.
Can you run Gemini 4 Neon inside Claude Code?
Yes. I ran Neon as a coding agent inside Claude Code through OpenRouter, because the Antigravity command line on our Google account offered no Neon. Claude Code is Anthropic's home field, so Google's own agent might get more out of Neon.
Should I switch from Claude Opus 5.5 to Gemini 4 Neon?
If you build games, 3D scenes, demos, or prototypes, Neon is worth it. If most of your day is client work and fixing bugs, I wouldn't switch just for this, because the quality was the same and Opus finished in less than half the time.
Chapters
- 0:00Neon built a neon city for $2.70. Opus gave me a black screen
- 0:35What Gemini 4 Neon is (price, context, Google's own table)
- 2:15Night Crew
- 2:30The independent check: Artificial Analysis, Vals, Arena
- 4:51What people built with Neon (and the doubts)
- 6:38How we tested: both in Claude Code, same prompt, one shot
- 7:51Test 1: three triple-A browser games
- 8:41Game 1: the city flight
- 10:05Game 2: the railgun robot
- 10:57Game 3: the coastal rally
- 12:15Test 2: a client app (12 hidden tests)
- 14:19Test 3: website design
- 16:44Test 4: the bug hunt (16 hidden bugs)
- 18:49The verdict: who should switch
Transcript
Read the full transcript
0:00Google's Gemini 4 Neon just built me a Neon City I can fly through in under 20 minutes for $2.70. Opus 5.5 got the same prompt, ran out of time, and handed me a black screen. And Google's own table says Neon beats Opus on 16 benchmarks and loses just two. But on real work, that gap disappeared. So, in this video, we're breaking down six head-to-head tests from AAA games to a client job to our hardest bug hunt. and if neon is worth switching to. I use Opus everyday to run my business. Let's jump right into it. So, what is Gemini 4 Neon exactly? Google DeepMind dropped it on November 18th. Google says it's adapted from Gemini 4 Carbon and it sits
0:45in the middle of the family. Carbon is for the hardest, longest projects. Hydrogen is the fast everyday one and neon, in Google's words, balances depth and efficiency. Logan Kilpatrick's launch post has over 1.4 million views. And he said it's punching above its model class in a meaningful way. It has a 1 million token context window and four thinking levels. Low, medium, high, and extra high. High is the default. Now, the price $3 in, $15 out per million tokens. Opus 5.5 is $4 in, $20 out. So, per token, Neon is 25% cheaper. And this is the part that got my attention. Look who Google put in its own launch table. Not Sonnet, Opus 5.5,
1:32the model a lot of us use every day for coding. 19 rows. Neon wins 16, Opus wins two, and one row has no opus score at all. And the two rows Opus wins, Terminal Bench 4.1 and Frontier Code 2.0, the two agentic coding rows, which is the closest thing in that whole table to the work I actually do. and Opus only wins them by about a point. Neon's biggest lead is a bug finding test called SWE sweep 39 versus 10, but Google ran that one with Neon inside Google's own agent and Opus inside Claude Code. So, not exactly apples to apples. These are Google's numbers, vendor reported. Take that for what you will. So, who actually checked them? Every night, Nightcrew researches what
2:17your market is talking about, writes the script, fact checkcks it, voices it in your voice, and animates it. In the morning, you approve it, and it posts to YouTube, Instagram, Tik Tok, and Facebook. Nightcrew. Nightcrew.studio. So, artificial analysis runs every model through the same set of tests with nobody's marketing team involved. On their intelligence index, Neon at its top setting scores 61. Opus 5.5 at its top setting scores 59. That puts Neon third overall and Opus fifth. And Neon is cheaper to run. $644 per task versus $104 for Opus, about 36% less. Now, one detail on that page matters a lot for our tests. High is Neon's default, and
3:03at high, Neon scores 60, one point under its max. Opus at high scores 52, and its max is 59. So, at the same setting name, Neon is basically running at full power and Opus isn't. Our bench ran both of them on high. Keep that in your back pocket because it comes back at the end. And Artificial Analysis has one more board that's closer to what we do, their coding agent index. That one tests the model and the agent together. Neon inside Google's anti-gravity scores 59. Opus inside Claude Code scores 58, one point. And Neon's cost per task was $6.86. 86 against $12.61 for Opus. So in its own agent, Neon matches Opus for about 46% less per
3:50task. Then there's Val. Neon 70.87, Opus 70.14. That's inside their error bars, so it's a tie. But Neon cost $7.40 per test and Opus cost $24.77. And Neon finished each test in about half an hour where Opus took about an hour and 25 minutes. And then the crowd. On launch day, Arena had Neon at number two on its agent leaderboard, one spot ahead of Opus. 2 days later, as of November 20th, Opus at its max setting has passed it. Opus 1336, Neon 1333, three points. And on Arena's webdev board, which is basically who builds the best web apps, Opus leads 1819 to 1765.
4:38So the independent picture is pretty interesting. Neon is about as smart, a lot cheaper per task, and the people voting on real sessions are calling it close to even. So what happens when you actually hand it real work? Before our tests, I wanted to see what people were actually building with it because that's where I got the idea for our first test. The one everybody saw was from the bugged dev. He gave Neon one image of a robotic dog with a rail gun on its back and asked it to build it in 3D. And it came back with a walk, a trot, and a gallop, legs that plant on the ground and recoil when it fires. 145,000 views. Absolutely insane. Then Kimismas flew through a futuristic city. Procedural towers, rain, flying traffic.
5:26133,000 views. Now, to be fair, that one was Neon's sibling, Gemini 4 Hydrogen, and it took him one follow-up. And people were racing, too. Camper AI GG had Neon crash a car through a concrete wall in pure 3.js. 45 minutes and about $7 inside anti-gravity. Those are his numbers. and a space racer got shared around 15 minutes and a$150 their numbers. But it's not all hype. Sh Neural with 178,000 views says Neon burns through his anti-gravity limits about three times faster. He measured a cash hit rate of 9% against about 95 for Opus. That's one person's measurement, but if you're on a usage plan, that's a big deal. And he's not the only one. One
6:13developer says you have to check its work more carefully than with Opus. and that it starts to hallucinate if you stay in one session too long. And another one bought a Google Pro plan just to test it and called it a good model, just overhyped and still not Opus 5.5 yet. His words. So I took three kinds of game people kept building a city flight, a robot with a rail gun, and a racer and turned them into our first test. So here's how we ran it. Both models ran as coding agents inside Claude Code. Same prompt, a fresh folder, high effort, web search off, one shot, and a 45minute limit, an hour for the bug hunt. Opus ran on my Claude subscription. So the cost you'll see for Opus is the API equivalent price. Neon ran through Open Router. So for Neon,
7:00that's what we actually got build. Now, why not Google's own agent? On our Google account, the only Gemini model the anti-gravity command line offers is Gemini 3.8 Flash. No Neon. You need a paid pro or ultra plan or an API key. So, I'll say it straight. Claude Code is Anthropic's home field and Google's own agent might get more out of Neon. And one more thing on settings, Opus normally runs on medium. We bumped it to high so both got the same setting name. And one more rule, if a model hasn't saved anything 15 minutes into a build, we restart it once in the same folder and the clock keeps running. Same rule for both. Remember that one. The scored tests are graded by hidden tests or scripted checks the models never see, plus two judge models from other labs
7:46that aren't in this matchup. A vibe check, not a verdict, but a pretty hard one. So, the prompts spell it out. AAA production values, real lighting and shadows, bloom, particles, physics, a HUD, and sound. Everything generated in code. No downloaded models. No downloaded textures. One HTML file that runs in the browser. These are our own prompts. The people on X didn't share theirs, and the bug dev used an image, so we wrote a text prompt for each idea that names every control and one must-have. A must-have is the one thing that makes it a real game and not just a pretty scene. Crash into a building and you respawn. Shoot a drone and it actually blows up. The rivals actually race you. And every game gets played
8:32with the exact same keyboard inputs frame by frame. So what you see is apples to apples. I didn't play these by hand and picked the best take for either model. Game one, the city flight, inspired by Kimonismos. You pilot a hoverjet through a huge futuristic city at night in the rain and you fly through glowing rings. And if you hit a building, you crash and respawn. Neon finished in 19.7 minutes for $2.70, and it never needed the restart. It saved a first version early and kept building on it. And look at this. Hundreds of towers with lit windows, neon signs, rain streaking past the camera, the rings light up, and when you clip a tower, crashed, and you're right back in at the last ring. That's a real game. Opus 15 minutes in, it hadn't
9:19saved a single file. It was still thinking, so it got the restart and then it hit the 45minute limit. The last version it saved runs. The ring counter works, the timer works, the speed and altitude work, but the city never draws. It's a black screen. Now, a run that ends with nothing playable on screen gets one retry under our rules. Same prompt, fresh folder. And on its second try, Opus saved its first version before the 15-minute mark and finished in 28 minutes. And this time, the city is there. Towers, rain, a purple sky, the crash, and the respawn. A good-looking game, but it's a second try. It costs $6.53 API equivalent, and it doesn't change the score. Neon did it on the
10:04first try. Game two, the rail gun robot, inspired by the bug devs demo. A four-legged robot dog with a rail gun on its back on a desert test range at Golden Hour shooting down drones. Neon took 17.4 4 minutes and $2.39. And the robot looks great. It feels mechanical. The legs plant. There's dust coming off the feet. The drones hover with their lights on. Opus needed the restart again. It finished at 37.8 minutes, and the API equivalent price was $4.84. Opus called its robot the Rail Dog Mark 2. And look at the hits. Big bloom on every explosion. Score popping off each target. Now look at the exact same 30 seconds of inputs. Opus' robot destroyed
10:50six targets. Neons destroyed one. Same keys, same timing. Which one looks better to you? Game three, the coastal rally. A race along cliffs above the ocean at sunset against three computer cars. Three laps. Neon finished in 13.9 minutes for $212. You've got the ocean, the palm trees, the lighthouse, the low sun, and the rivals actually drive the track by themselves. Lap counter, position, mini map, all there. Opus needed the restart one more time, and it ran right into the 45minute limit. But the version it saved plays a real asphalt road along the cliffs. Guard rails, tire smoke when you drift, the rivals pulling away, and a mini map. Now, to be fair, our scripted
11:37steering is blind. So, in both versions, our car spends a lot of time scraping the rail. That's on our inputs, not the models. So, the game test goes to Neon. Three playable games out of three. Everyone under 20 minutes, about $7 for all three. Opus had to be restarted on all three. And on its first tries, it finished one inside the time limit. And like I've said before, Opus' nickname is Slopus. It has been slow since the dawn of time. But when Opus did finish, the robot game was the better game to actually play. Comment down below, which of these six builds would you actually play? So, games are fun, but would it do my real work? Test two is a realfeing freelance job. A two partner accounting
12:22firm, Grace and Daniel, wants a page where clients upload their tax documents, and they send it as a messy email thread. There's a trap in that thread. Grace says everyone sends a photo ID and then partway through Daniel changes it. Returning clients skip the photo ID and last year's tax return and there's a question nobody answers which is what to do with a married couple filing together. The model has to build the app and write the email back to the client. Then 12 hidden tests check every requirement including the one that changed. You can see the app here. The client types their name and email ticks I was a client last year. picks what applies to them and gets a checklist built for them. They upload each document. The progress bar fills up and they submit. Then the firm flips over to
13:08a staff dashboard with every client, their status, and how many documents they're still missing. Neon 12 out of 12, 10 minutes and 40 seconds, $1.56. Opus, 12 out of 12, 4 minutes and 36 seconds, and the API equivalent was $1.24. both caught Daniel's change. And on the married couple question, they came up with the exact same answer. Let the W2 item take more than one file so both spouses forms can go in the same spot and then ask Grace and Daniel to decide if they want separate spouse items later. Same idea, and each one got there on its own. And both emails were honest about the limits. both told the client this page doesn't actually send files to the firm yet and that real file storage and a staff login are the next step which is exactly what you want a
13:54developer to tell a client instead of overpromising and the reply emails were scored by two judges from labs that aren't in this test v4 pro and Quen 3.8 8 Max Prime. Both models got 10 out of 10 from both judges. So, would it do my real work? Yes, both would. Opus just did it in less than half the time. For a client job, that's the difference between waiting on it and moving on to the next thing. Test three is brand new, a website design job. I'm tired of seeing AI slop websites out there, so I wanted to see who actually has taste. The client is Spoke and Mend, a bike repair shop on Alberta Street in Portland, run by two mechanics. One page, one file, nothing loaded from the internet, and it has to work on a
14:40desktop and on a phone. No design skill or plug-in for either model, just their own taste. We judged it three ways. First, hard checks by script. Nothing runs off the side on a phone. Every bit of text has a contrast of at least 4 1/2 to one against its background. Every section the brief asks for is there. The book button shows up without scrolling and zero network requests. Neon passed all five. Opus passed four. Nine bits of its orange text missed it. Most came in at 4.35 to1, just under the line, and the worst at 3.83. Neon took 11 12 minutes and $147. Opus took about 8 minutes and the API equivalent was $229. Second, two vision judges from other
15:25labs, Mistral Large 4 and Kimmy K 3.1 Pro. They got full page screenshots at both widths with the names hidden and scored six things: hierarchy, type, spacing, color, originality, and would you trust this shop? And both judges picked opus 111 to 109 out of 120. Two points. They loved its repair tag idea. The prices hang off a shop wire like real repair tags, and the booking form looks like a tag, too. And they weren't perfect. One judge said Opus' dropoff date strip gets cut off on a phone, and that neon service rows wrap awkwardly on a phone with some tiny faint labels. And here's what made me laugh. Look at the colors. Both models picked almost the exact same palette. Cream, forest green,
16:11and a rusty orange. Then I asked each one to write a few lines on its design direction. Opus said it built the whole page around the paper repair tag that gets tied to every bike in a real shop. Neon said it drew from printed rate boards and punched Manila repair tags. They both landed on the repair tag. Same prompt, same instincts. Neon just put it on a booking ticket and Opus put it on every single price. Third, you here they are side by side on a desktop and on a phone. I'm not a designer, but which one would you book your repair from? Test four is the extreme one. Tally house, an invoicing backend. About 3,000 lines of code across 42 files. We hid 16 real bugs in it. Most of them spread across
16:57two files, plus 11 decoys that look like bugs, but aren't. The models aren't told how many there are. Hidden tests decide what got fixed, and two judges decide what got explained. Breaking something or reporting a fake bug costs points. 32 is perfect. These are bugs that would actually hurt a business. An API key from one company that can read another company's invoices. A customer getting charged twice when two payments land at the same time. A date that shifts when daylight saving time kicks in. Tax rounding that's off by a cent. Opus 8.2 minutes, $2.51 API equivalent. 14 of 16 fixed and explained. Nothing broken. No fake bugs. 28 out of 32. Neon, 17.4 minutes, $2.92.
17:4814 of 16 fixed and explained. Nothing broken, no fake bugs. 28 out of 32. The exact same score. And neither one fell for a single decoy. 11 traps built to look like classic bugs. And both models walked right past everyone. And both judges, Kimmy K3.1 Pro and GLM 5.3 Prime, agreed on both, and they missed the exact same two bugs. One is a report cache that gets filled with old numbers when a payment lands at the wrong moment. The other is a refund that never turns the payment reminders back on. Both of those only show up when you think about what happens over time, not by reading one line. Neon also scored 28 on this exact hunt two days ago. So that
18:34number is consistent. And to be honest, one model has hit a perfect 32 on it from one of its two judges, GPT 6.2 Soul. So the top of this test is getting crowded. Same score here. Opus just got there in half the time. So let's put it all together. Across all six tests, Neon's bill came to $13.15. On the games, Neon wins. Three for three, under 20 minutes each, about $7 total. On the client job, a tie, 12 for 12 and 10 out of 10. On the design, it's split. Neon passed every hard check, and the judges liked Opus by two points. And on the bug hunt, a tie at 28, so Google's table said 16-2. On our tests, the quality was basically even. The real
19:20difference was speed, and it flips depending on the job. On the big game builds, Opus sat there thinking past the 15minute mark on all three scored runs. On the client job and the bug hunt, Opus finished in less than half the time. And on the design, it was still about 30% quicker. And remember the settings. We ran both on high. For Neon, that's basically its max. For Opus, artificial analysis says max is 7 points above high. So, this was neon close to its best against Opus with something left in the tank. So, who should switch? If you build visual stuff, games, 3D scenes, demos, prototypes, Neon is the real deal and it's cheap per token. If you're already paying for Google AI Pro or Ultra and you live in anti-gravity, it's
20:05worth a real try. Just watch your limits. And if most of your day is client work and fixing bugs, I wouldn't switch just for this. The quality was the same, and Opus got there faster. For me, Opus is still my daily driver. I'm on the $200 plan, so I'm not paying per token. and the stuff that pays my bills looks a lot more like that client job than a neon city. Think a quick tool for the reselling side of my business. On the client job and the bug hunt, Opus did the same quality work in less than half the time. But man, that neon city. Wow. So, after six tests, I want your call. Are you moving to Neon, staying on Opus, or splitting the work between them? Drop it in the comments. We'll see you in the next one. Have a blessed day.
Ready to start building with AI?
Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.
Join Shipping SkoolKeep reading
Gemini 4 Carbon vs Claude Opus 5.5: Benchmarks and Speed Test
Gemini 4 Carbon tied Claude Opus 5.5 on both of my hidden-test builds (12/12 and 28/32 each), but in my runs through OpenRouter it took up to 7.6 times longer and twice saved no file. Carbon costs $2 in / $10 out until November 30, then the same as Opus.
GPT-6.1 Terra vs Claude Sonnet 5.5: Coding Tests, Cost, and Speed
GPT-6.1 Terra costs half as much as Claude Sonnet 5.5 and finished each of my tests in about a minute. It tied Sonnet on a hard bug hunt, but Sonnet clearly won both visual builds. Use Terra for fast, cheap everyday jobs and computer use, and Sonnet when the result has to look great.
GPT-6.2 Astra vs Sol vs Claude Fable 5.2: 4 Coding Tests and Costs
I ran GPT-6.2 Astra and Sol in Codex against Claude Fable 5.2 and Sonnet 5.5 in Claude Code on four tests. Sol won the bug hunt for 89 cents and cost $2.43 across all four tests, versus $9.14 for Astra and $35.88 for Fable (API equivalent).
