AI Coding Subscriptions: Where to Go After GitHub Copilot Changes

// Detect dark theme var iframe = document.getElementById('tweet-2051218167780041147-163'); if (document.body.className.includes('dark-theme')) { iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2051218167780041147&theme=dark" } This is how much it costs Github to run its Copilot. The era of request-count-based pricing has come to an end. Github is moving to token-based pricing. The cheapest plan was $10/month and it gave you 300 premium requests to advanced models. One premium request included all the inner back-and-forth of an agent until you were prompted to respond to the agent. In practice, this meant one request was more like 5-10 requests.

Let's see how the new pricing compares on the market.

This is a developing story.

On April 27, GitHub announced its move to a token-based pricing model. On May 12, the company introduced "flexible allotment" to its token based pricing, after receiving criticism for downgrading its previous service.

As of now, the prices themselves have not changed. Except for abolishing annual plans, GitHub says the monthly subscription tiers are still priced like before. What changed is how much you can get from the price you paid.

This total usage is deducted each time you use their API based on the pricing for each model, which are set to market rates. There are no more free models. The only thing free remaining is code completion. There is also no mention of continuing the "auto" model selection discount, which currently grants a 10% discount to users.

It is not clear how the the "flex" part of the pricing is determined. In their May 12 announcement, GitHub says this amount is variable.

Any leftover usage expires at the end of the month.

Existing annual plan holders will be grandfathered with their request-based pricing until their plan expires, but the model consumption multiplier is set to increase by up to 900%, meaning subscribers could get their usage reduced to 10% of what they enjoyed previously.

Model selection is also limited by the tier of subscription. For example, Claude Opus is not available on Pro plans.

Staying within the Microsoft umbrella, let's take a look at how competitive Codex subscriptions are.

| Plan | Price (Monthly) | GPT-5.5 (5h window) | GPT-5.4 (5h window) |
|---|---|---|---|

| Free | $0 | unknown | unknown | | Go | $8 | unknown | unknown | | Plus | $20 | 15–80 messages | 20–100 messages | | Pro | $100 | 80–400 messages | 100–500 messages |

OpenAI is not transparent about the usage limits for 5h and weekly windows. They only give a range of messages. From my experience, I find letting GPT-5.5 vibe code on its own for 30 minutes consumes about 25% of my 5h limit and 5% of my weekly limit.

One limitation of Codex is the model selection being restricted to OpenAI GPT models.

OpenCode is a coding agent by Anomaly. It is a terminal UI application. Here are the tiers of their subscription offering.

OpenCode's free model is Big Pickle, which is a stealth model. The paid models are open-source models like GLM, Qwen and DeepSeek. To put things in perspective, $12 of compute is about 880 calls to GLM 5.1, 3,300 calls to Qwen 3.6 Plus, and 31,650 calls to DeepSeek V4 Flash.

Google offers Gemini Code Assist for free to individuals. This service is available as a plugin inside IDEs and as a CLI. Only Gemini models are available. Although it is free, a Google AI subscription can increase the usage limit.

Let's not forget about the OG agent. Claude Code is limited to Anthropic models and it does not have a transparent usage limit. There is a peak hour pricing during which the 5h window is consumed faster.

Cursor is a VS Code fork with integrated AI coding features. Its model selection spans across all the biggest providers, in addition to an in-house model called "Composer".

Similar to GitHub Copilot, Cursor has moved to token-based pricing with an unclear included usage. Each model has its own cost per million token. Leftover credits/usage will not roll over to the next month.

Windsurf is another VS Code fork with AI integration. As of March 19, 2026, Windsurf has a variable usage limit pricing for a monthly subscription. The variable usage depends on the complexity of each message.

| Pro | $20 | 8-101 messages (Sonnet 4.6) | Full |
| Max | $200 | 47-631 messages (Sonnet 4.6) | Full |

Windsurf's model selection includes all the flagship proprietary models and the leading open-source models, for paid subscribers. In addition, Windsurf has an in-house model called SWE.

On top of a daily usage limit, there is also a weekly usage limit. After exceeding those limits, you will be charged based on API token pricing.

Let's wrap this up.

Even though I didn't mention this explicitly in each case, every service listed here has a free plan. But most of them do not have transparent usage limits to be judged fairly. Here are the services with clear usage limits.

The best service here looks like Gemini.

For this tier, we'll look at everything priced at and below $20 per month.

The most economic and reliable choice in this tier is OpenCode Go at $10/month. The most generous tier seems tied between Codex Plus and Claude Code.

Not every provider has this tier.

The best service here is GitHub Copilot Pro+.

Here are the tiers with prices starting at $100 per month.

For the $100 price, the best service is tied between Claude Code Max 5x and Code Pro. The choice depends your preference between Claude and GPT.

For $200, the best service is Claude Code Max 20x if you use it on a consistent basis. However, if you use AI in heavy bursts, Cursor Ultra is the better choice.