Skip to main content
TokenCost logoTokenCost
Model ReleaseNovember 21, 2026·7 min read

MAI Code 2 costs 41% of GPT-6.1 Sol per task, at a half-price promo. On May 17 it becomes GPT-6.1 Terra's card to the cent, and Flash ends up costing 30% more than Luna.

Microsoft's cost claims are real numbers from a real chart. They are also all measured at half price, and the footnote saying so is easy to miss.

Close-up of colorful minified JavaScript code on a black screen, angled and fading into darkness

Photo by Mitchell Luo on Unsplash

Microsoft shipped MAI Code 2 and MAI Code 2 Flash at Ignite on November 17. GitHub's Copilot pricing page lists Code 2 at $0.50 per million input tokens, $0.025 cached and $2.50 output, and Flash at $0.05 / $0.005 / $0.25, with no long-context step and no cache write fee. The model card gives both a 1,050,000-token context.

The headline from Microsoft's launch post is that Code 2 delivers 98% of GPT-6.1 Sol's performance at 40% of the cost, and 96% of Claude Opus 5.5's at 5%. A footnote adds that MAI costs are shown at a 50% promotional price. The Copilot page dates it: promo rates run until May 17, 2027, and every line doubles after that.

Doubled, the two cards stop being new. Code 2 becomes $1 / $0.05 / $5, which is GPT-6.1 Terra's card to the cent. Flash becomes $0.10 / $0.01 / $0.50, which is GPT-6 Luna's and Claude Haiku 5.5's. Microsoft is renting you six months at half of OpenAI's mid and budget prices.

  • Microsoft's 40%-of-Sol and 5%-of-Opus figures check out on its own chart, at half price
  • after May 17, 2027 Code 2 is roughly 81% of Sol per task and Flash costs about 30% more than Luna
  • right now you can only buy Code 2 through GitHub Copilot

Both cards, before and after May 17

Per 1M tokensCode 2 nowCode 2 from May 17Flash nowFlash from May 17
Input$0.50$1.00$0.05$0.10
Cached input$0.025$0.05$0.005$0.01
Output$2.50$5.00$0.25$0.50
Same card as-GPT-6.1 Terra-GPT-6 Luna, Haiku 5.5

GitHub Copilot models and pricing, read November 21, 2026. No cache write column and no long-context tier on either model.

Where you can actually buy this matters. Right now it's GitHub Copilot on usage-based billing, where 1 AI credit is $0.01. Code 2 is on Pro, Pro+, Max, Business and Enterprise. Flash is on those plus Free and Student. Annual plans still on request billing see a multiplier of 1 for Code 2 and 0.33 for Flash, per the multiplier table.

Flash is also in a limited Microsoft Foundry preview at the same $0.05 / $0.25 (US Gov regions are 1.375x). Code 2 has no Foundry meter yet. OpenRouter, Vercel and Kilo are "coming soon" per Microsoft, and the model card says weights aren't distributed.

Microsoft's own numbers, at both prices

Microsoft averaged four benchmarks for Code 2 (DeepSWE 1.1, Terminal-Bench 2.1, SWE-Atlas, FrontierCode 1.1) and three for Flash, every model at max effort through the Copilot CLI with subagents off. The per-task costs come from its launch chart. The last column is ours: MAI's cost doubled, everyone else's unchanged.

Model (max effort)Avg score$ / task now$ / task from May 17, 2027
MAI Code 263.2$0.309$0.618
GPT-6.1 Sol64.1$0.761$0.761
Claude Opus 5.565.3$6.306$6.306
Model (max effort)Avg score$ / task now$ / task from May 17, 2027
MAI Code 2 Flash61.0$0.052$0.104
GPT-6 Luna55.1$0.080$0.080
Claude Haiku 5.553.9$0.319$0.319

Scores and costs from Microsoft's chart as reported by InfoWorld; the MAI DeepSWE, Terminal-Bench and SWE-Atlas scores match the model card. Code 2's average covers four benchmarks, Flash's three. The May 17 column is our arithmetic.

The arithmetic checks out at today's price. $0.309 against $0.761 is 41% of Sol, and against $6.306 it's 4.9% of Opus. From May 17 the same task costs $0.618: 81% of Sol and 9.8% of Opus. Still a big gap to Opus. Against Sol it mostly closes, and Sol scores a point higher.

Flash flips. At $0.052 it's 65% of Luna today, with a six-point higher average. At $0.104 it costs 30% more than Luna per task. Whether that's worth it depends on how much you value those six points, which come mostly from Terminal-Bench 2.1 (78.3 vs 69.3) and SWE-Atlas (44.8 vs 36.4).

One claim doesn't survive the chart. Microsoft's blog says Flash runs at 20% of Haiku 5.5's cost. $0.052 over $0.319 is 16%. That's Microsoft underselling itself, which is rare enough that we'll mention it.

The tokens behind the per-task numbers

Nobody publishes output tokens per task for these models yet, and Artificial Analysis has no listing for either. But you can back them out. Code 2's card is a quarter of Sol's, yet its task costs 41% of Sol's, so it spends roughly 1.6 times Sol's tokens on the same work. Flash spends about 1.3 times Luna's. Opus spends about 2.5 times Code 2's.

That ratio is why the promo matters so much. A cheaper token only beats a pricier one if you don't use too many more of them. Code 2 uses 1.6x Sol's tokens at a quarter of the price today, and at half of Sol's price after May. Flash at 1.3x Luna's tokens and the same price as Luna has no margin left.

Two caveats on the inputs. These ratios assume similar input and output mixes across models, which is a guess. And the FrontierCode figures in Microsoft's chart don't match Artificial Analysis's live table, which Microsoft cites as the source: AA shows Opus 5.5 at 52.2 where the chart says 56, and Sol at 49.1 against 49.4. MAI Code 2 isn't on that table at all, so its 52 can't be checked.

A review bot's monthly bill

Per-task numbers bake in each model's verbosity. Sometimes you just want the meter for a job you already run. Say a bot reviews 200 pull requests a day: a 30,000-token diff with nothing cached, and a 3,000-token review back. Over 30 days that's 6,000 calls.

MAI Code 2 Flash (promo)
$13.50
Flash from May 17 / Luna / Haiku 5.5
$27.00
DeepSeek V4.1 Flash (off-peak)
$37.80
MAI-Code-1.1-Flash
$57.60
MAI Code 2 (promo)
$135.00
Code 2 from May 17 / GPT-6.1 Terra
$270.00
GPT-6.2 Sol
$540.00
Claude Opus 5.5
$1,080.00

Our arithmetic from the GitHub, OpenAI and Anthropic price pages, same token counts for every model. Haiku 5.5 stays on its short-prompt card below 100K tokens.

Code 2 runs that bot for $135 a month today and $270 after May 17, which is exactly the Terra bill. GPT-6.2 Sol is $540, Opus 5.5 $1,080. Flash does it for $13.50 until May and then ties Luna and Haiku 5.5 at $27. Keep in mind this chart charges every model the same tokens, and the section above says Code 2 and Flash don't use the same tokens.

Long prompts are where Code 2 keeps an edge after the promo. OpenAI bills the whole request at a higher rate past 272K input tokens. Feed in a 400,000-token slice of a monorepo, uncached, and ask for 6,000 tokens back: GPT-6.2 Sol charges $1.69 and Terra $0.845. Code 2 charges $0.215 now and $0.43 after May. Flash is even more interesting here. At $0.043 after the promo it undercuts Luna's $0.0845, because Luna steps up past 272K and Flash doesn't.

The fine print, and where it disagrees with itself

Microsoft's launch post says both models are not yet tuned for very long-horizon agent work. The model card says the opposite, recommending Code 2 for the most complex long-horizon tasks. The card also says they're optimized for English only, rates their alignment as similar to GPT-5.6 Sol and below GPT-6, and puts the pre-training cutoff at August 2026.

Then the card contradicts itself. Its summary table lists Code 2 at 80B active parameters, while the body text says 77B active out of about 2.8 trillion. And a second "advanced autonomy" table gives Code 2 60.2% on DeepSWE 1.1, against 73.2% in the main table, with no explanation of the different setup. Same benchmark, 13 points apart.

Flash also beats Code 2 on SWE-Atlas in the main table, 44.8 to 43.1. If your work looks like SWE-Atlas, the cheap one might be the right one.

Who should switch before May

If you already pay for Copilot, try Code 2 now. For six months it's about 40% of Sol per task on Microsoft's numbers, and even at full price it should stay cheaper than Sol and far cheaper than Opus. Just don't build a budget on the promo number. Plan for $1 / $5 and treat anything below that as a bonus.

Flash is a harder call. It's the best deal in budget coding models until May 17. After that it's Luna's card with more tokens per task, so it has to earn its keep on scores alone. Run it against Luna on your own tasks before then and see whether the gap Microsoft reports shows up.

If you work through the raw API rather than Copilot, there's nothing to decide yet for Code 2. It isn't sold anywhere else today.

We covered Microsoft's first coding model in MAI-Code-1-Flash, and the cards it now copies in GPT-6.1 Terra, GPT-6 Luna and Claude Haiku 5.5. For the newer OpenAI baseline see GPT-6.2 Astra and Sol, or run your own numbers in the calculator.

Sources