GitHub Copilot's New Multiplier Table: 5 Models That Just Got Dramatically More Expensive

GitHub Copilot Just Revealed How Deep Its Subsidies Were — And the Numbers Are Stark

On June 1, GitHub Copilot’s new multiplier table takes effect. Claude Opus 4.7 jumps from 7.5x to 27x. Gemini 3.1 Pro and GPT-5.3 Codex both jump from 1x to 6x. If you’re a Copilot subscriber who has been running agentic coding sessions on frontier models, your effective cost is about to increase by a factor of 3.6 to 6 depending on which model you’ve been leaning on.

That’s not a price hike in the normal sense. That’s Microsoft revealing, in public, exactly how much it had been absorbing on your behalf.

GitHub CPO Mario Rodriguez explained the shift plainly: “Today, a quick chat question in a multi-hour autonomous coding session can cost the user the same amount. GitHub has absorbed much of the escalating inference cost behind that usage, but the current premium request model is no longer sustainable.” The new model mirrors Cursor’s approach — a monthly credit allotment, with the option to buy more. The preview period runs through May so you can see your projected bill before the switch flips.

Here are the five models hit hardest by the new table, what the numbers actually mean, and what you should do before June 1.

---

Claude Opus 4.7: From 7.5x to 27x

The single most dramatic change in the table. Opus 4.7 goes from a 7.5x multiplier to 27x — a 3.6x increase in effective cost per session.

To understand what this means in practice: if you were running a multi-hour autonomous coding session with Opus 4.7 under the old model, you were paying roughly $0.39 per credit-unit of compute. Under the new model, that same session costs $1.40 per credit-unit. For power users running long agentic loops — the exact use case Copilot has been marketing — this is a material budget line.

The 27x multiplier also tells you something about Anthropic’s underlying API pricing. Opus 4.7 is priced at $5 per million input tokens and $25 per million output tokens. Agentic sessions are output-heavy. A multi-hour coding session that touches multiple files, iterates on tests, and reads back its own diffs can easily generate hundreds of thousands of output tokens. At $25/million, that adds up fast — and Microsoft had been eating the difference.

Developer Peter Deneen put it bluntly: “These new Copilot multipliers starting June 1st are absolutely ridiculous. I can only imagine this pricing is going to force users to lock in with a single foundation model vendor just to manage costs.” That’s a reasonable read. When Opus costs 4.5x more than the next tier down in the same tool, you start making different architectural decisions.

For teams that have built workflows specifically around Opus 4.7’s reasoning quality — complex refactors, multi-file architecture work, anything requiring sustained context — this is the model to audit first. The question isn’t whether Opus is worth it in absolute terms. The question is whether it’s worth 4.5x more than a capable alternative for every step of your pipeline. For most workflows, the answer is no.

If you want a direct comparison of where Opus 4.7 actually outperforms cheaper alternatives on real coding tasks, GPT-5.5 vs Claude Opus 4.7 real-world coding performance has the specifics.

---

GPT-5.3 Codex: From 1x to 6x

GPT-5.3 Codex going from 1x to 6x is the change that should concern the largest number of Copilot users, because 1x was essentially “free within your subscription.” A lot of teams defaulted to Codex precisely because it felt like the included option.

It wasn’t free. Microsoft was just absorbing the cost.

The 6x new multiplier means Codex now costs the same as Gemini 3.1 Pro in the new table. That’s a deliberate signal: the era of one model being the obvious cheap default inside Copilot is over. Every model now has a real cost attached to it, and the spread between cheap and expensive has compressed somewhat while the floor has risen significantly.

OpenAI’s Codex has been growing fast — 200,000 users on January 1 to 4 million the week before GPT-5.5 launched, a 20x increase in roughly four months. That growth is exactly why the 1x pricing was unsustainable. You can’t absorb frontier model inference costs for 4 million active users on a flat subscription.

For teams that have been using Codex as their workhorse model inside Copilot, the practical question is whether GPT-5.4 Mini or another sub-frontier model can handle the routine parts of your pipeline. GPT-5.4 Mini vs Claude Haiku 4.5 for sub-agent use cases is worth reading before you make that call — the benchmark gaps between frontier and sub-frontier models have narrowed considerably for well-defined tasks.

---

Gemini 3.1 Pro: From 1x to 6x

Gemini 3.1 Pro gets the same treatment as Codex — 1x to 6x. This one is interesting because Gemini has been the underdog choice inside Copilot, often selected by teams that wanted Google’s long-context capabilities or were already in the Google ecosystem.

At 1x, it was a compelling option for context-heavy tasks: large codebase analysis, documentation generation, anything requiring a wide context window. At 6x, the calculus changes. You’re now paying the same rate as Codex for a model that, in most coding benchmarks, trails both GPT-5.3 and Opus 4.7.

Gemini’s pricing on the direct API is $2 per million input tokens and $12 per million output tokens — cheaper than Opus but more expensive than some alternatives. The 6x Copilot multiplier doesn’t change the underlying API economics, but it does mean that the Copilot abstraction layer is no longer hiding those costs from you.

For teams using Gemini 3.1 Pro specifically for its context window, this is a good moment to test whether DeepSeek V4 — which offers a 1 million token context window at $1.74 per million input tokens and $3.48 per million output tokens — can handle the same workload at a fraction of the cost. It’s open-weight, which means you can run it on your own infrastructure if data residency matters to you.

---

The Models That Didn’t Move (And What That Tells You)

The multiplier table changes aren’t uniform. Some models stayed flat or moved less dramatically. That’s deliberate.

GitHub is using the new pricing structure to steer behavior. The models that stayed cheaper are the ones Microsoft wants you to use for routine tasks — the sub-frontier models that cost less to serve and where the performance gap for most coding tasks is smaller than people assume. The models that jumped are the ones that were being used for everything, including tasks that didn’t require their full capability.

This is the same logic that Replit applied when it moved to usage-based pricing in the summer and fall of 2025 — earlier than most platforms, and it took real criticism for it. The lesson Replit learned, and that GitHub is now applying, is that flat-fee pricing for agentic workloads creates a tragedy of the commons. Power users consume disproportionately, the platform degrades for everyone, and the only fix is either rate limits or honest pricing.

GitHub chose honest pricing. The multiplier table is the mechanism.

If you’re building multi-model pipelines and want to route tasks to the right model at the right cost, platforms like MindStudio make this orchestration tractable — 200+ models, visual workflow builder, and the ability to swap models per step without rewriting your integration layer. The model routing problem the new Copilot table creates is exactly the kind of problem that benefits from a layer of abstraction above any single vendor’s pricing.

---

The Hidden Model: What the Old Table Was Really Pricing

There’s a fifth entry in this story that doesn’t appear on the multiplier table directly: the implicit subsidy model that GitHub was running until now.

When Opus 4.7 was at 7.5x, that wasn’t the real cost. The real cost, revealed by the jump to 27x, was approximately 3.6x higher. Microsoft was absorbing that difference — not as a loss leader in the traditional sense, but as a strategic bet that agentic coding adoption would grow fast enough to justify the subsidy. It did grow. And then it grew too fast.

GitHub’s stability was compromised by excess traffic. Their own blog post acknowledged it: “Usage-based billing fixes that. It better aligns pricing with actual usage, helps us maintain long-term service reliability, and reduces the need to gate heavy users.”

The old table wasn’t just underpriced. It was structurally dishonest about what AI inference actually costs. Every team that built their AI cost models around Copilot’s old multipliers now has a gap between their projections and reality. Goldman Sachs recently reported that companies are blowing past AI inference budgets by orders of magnitude, with inference costs in engineering approaching 10% of total headcount costs. The Copilot repricing is one data point in a broader pattern.

Boris Cherny at Anthropic said it plainly when explaining why Anthropic was pushing heavy users toward the API: “Subscriptions weren’t built for the usage patterns of these third-party tools.” That’s true of Copilot’s old model too. A $39/month subscription wasn’t built for multi-hour autonomous coding sessions consuming millions of tokens.

---

What to Do Before June 1

The preview period is the useful part. GitHub is showing you your projected bill under the new model before the switch happens. Use that data.

First, pull your Copilot usage breakdown by model. If Opus 4.7 is your default for everything, you’re about to see a significant increase. The question is whether you need Opus for everything, or whether you’ve just never had a reason to route differently.

Second, run a model comparison on your most common task types. For well-defined coding tasks — unit test generation, boilerplate, documentation — sub-frontier models often match frontier performance. How to save tokens in Claude Code using Opus plan mode covers one specific technique: plan with Opus, execute with Sonnet. The same logic applies inside Copilot — use the expensive model for the reasoning step, not the execution step.

Third, if your team has been using Copilot as a cost-transparent abstraction layer, that abstraction is now leaky. You need visibility into which models are being called, how often, and for what task types. This is the “AI cost scoreboard” step — making agent economics visible to the people making architectural decisions.

The broader question the new multiplier table forces is whether Copilot is still the right abstraction for your team’s agentic workflows. For some teams, the answer is yes — the integration with VS Code and GitHub Actions is worth the premium. For others, the new pricing makes direct API access with your own routing logic more attractive. How to use Open Router free models with Claude Code to cut AI costs is one path if you want to move off the Copilot abstraction entirely.

For teams building production applications on top of these models — not just using them for coding assistance — the spec-driven approach is worth understanding. Remy compiles annotated markdown specs into complete TypeScript stacks: backend, database, auth, deployment. The source of truth is the spec; the generated code is derived output. When model pricing shifts, you update your routing logic in the spec, not across a codebase.

---

The Multiplier Table as a Signal

The most important thing the new Copilot multiplier table tells you isn’t the specific numbers. It’s that the numbers were always there — Microsoft was just hiding them from you.

Every AI platform that has been running flat-fee pricing for agentic workloads has been making the same bet: that usage would stay manageable, or that growth would eventually justify the subsidy. For GitHub, neither held. The agentic era arrived faster than the pricing model could absorb.

My opinion: this is actually good news for teams that want to build durable AI workflows. Honest pricing forces honest architecture. When Opus 4.7 costs 4.5x more than the next tier, you stop using it as a default and start using it where it actually matters. That’s better engineering, not worse.

The teams that will be hurt most by June 1 are the ones that haven’t looked at their model usage in months. The teams that will be fine are the ones that treat model selection as an ongoing decision, not a one-time default.

For a side-by-side look at where the frontier models actually differ on coding tasks — which matters a lot when you’re deciding whether the 27x multiplier is worth it — GPT-5.4 vs Claude Opus 4.6 vs Gemini 3.1 Pro benchmark results has the data. The performance gaps are real but narrower than the pricing gaps suggest.

The subsidy era is over. The multiplier table is the receipt.