AI Coding Tools Pricing Shift 2026: Flat-Fee Era Ends, Token Billing Begins
Between 20 and 23 April 2026, the largest AI coding tool providers reshaped their pricing. Microsoft is moving GitHub Copilot to token-based billing, Anthropic tests Claude Code restrictions in the Pro plan, GitHub has paused individual signups entirely. European teams need to rebuild their budget and AI tool policies before the new tariffs become the norm.
Microsoft is moving GitHub Copilot to token-based billing from June 2026: Business at 19 USD per user per month with 30 USD in AI credits, Enterprise at 39 USD with 70 USD in credits. On 21 April 2026 Anthropic briefly removed Claude Code from the 20 USD Pro plan, reversing the change after backlash but continuing a test on 2 percent of new prosumer signups. GitHub paused new registrations for Copilot Pro, Pro+ and Student plans on 20 April 2026, citing weekly compute costs that nearly doubled since January 2026. Flat-fee subscriptions miss the actual token cost for heavy users by up to a factor of 10. For European companies budgeting becomes harder in 2026: per-developer token billing can trigger works council co-determination under Section 87 BetrVG, Opus 4.7 costs 5 USD per million input tokens and 25 USD per million output tokens. Recommendation: introduce consumption transparency, set runtime guards, involve the works council early and evaluate open-source alternatives like Continue.dev, Ollama and Kimi K2.6 as a hedge.
Context: 72 hours reshape the cost logic
Between 20 and 23 April 2026 the two biggest AI coding tool providers overhauled their pricing. The simultaneity is not coincidence, it is signal: flat-fee subscriptions no longer cover the real cost of agentic workflows. European engineering teams using Copilot or Claude Code face a new budget logic.
The backdrop is sober: long-running agent sessions consume many times what a pauschal subscription could absorb. GitHub speaks of fundamentally changed compute demands, Anthropic of sharply rising engagement intensity. Both companies are therefore shifting the economic load to the user, or more precisely to the employer's engineering budget.
Agentic workflows have fundamentally changed Copilot's compute demands. Long-running, parallelized sessions now regularly consume far more resources than the original plan structure was built to support.
For innobu's audience, strategic decision makers in mid-market and enterprise, this is an early indicator. The same logic will hit every enterprise AI offering sooner or later: the subsidy phase ends, usage-based billing becomes the standard. Companies that install measurement points now act prepared rather than reactive. Background context in our AI Coding Tools Comparison 2025 and in our analysis of Claude Code Routines in the Cloud .
What Microsoft and Anthropic are changing
The three events followed in quick succession. Each was notable on its own, together they mark the end of the flat-fee era for AI coding tools. The timeline shows how fast the market has turned in a matter of days.
GitHub pauses new signups for Copilot Pro, Pro+ and Student
Joe Binder, VP of Product, cites weekly compute costs nearly doubling since January 2026. At the same time Claude Opus 4.5 and 4.6 are removed from Pro+, Opus 4.7 remains with a 7.5x request multiplier until 30 April.
Anthropic removes Claude Code from the Pro plan and reverses the move
The pricing page briefly shows an X instead of a checkmark next to Claude Code in the 20 USD tier. After backlash Anthropic reverses the change the same day. Head of Growth Amol Avasare confirms a test on 2 percent of new prosumer signups.
The Register and wheresyoured.at document the scope
Trade media report that flat-fee subscriptions miss actual token costs by up to a factor of 10 for heavy users. A Max 20x user can burn through the equivalent of 600 to 1,500 USD in API tokens per month for a 200 USD flat fee.
Microsoft announces token-based billing from June 2026
Business pays 19 USD per user per month with 30 USD in AI credits, Enterprise 39 USD with 70 USD in credits. Claude Opus 4.7 is billed at 5 USD per million input and 25 USD per million output tokens. Google reports in parallel that 75 percent of new code is AI-generated.
Anthropic acknowledges Claude Code quality issues
The provider confirms three separate error sources that made Claude Code perform worse for weeks. Under the new token billing model users will directly pay for failed agent runs.
Why the simultaneity matters: When GitHub, Microsoft and Anthropic move in the same direction independently within four days, that is not a vendor problem, it is a structural signal. The entire industry is repricing. Any company that built its budget on flat-fee assumptions should remodel it before June.
Comparison: Old flat fee versus new token billing
The new Copilot tariffs look stable at first glance, but the pooled AI credits mask the real cost dynamics. A team with active agent usage can burn through its credits in days and pay overage for the rest of the month. The table below shows the key metrics of the new structure.
| Plan | Before (pre April 2026) | New (from June 2026) |
|---|---|---|
| Copilot Pro (Individual) | 10 USD per month, 300 premium requests | New signups paused, Opus models removed |
| Copilot Pro+ (Individual) | 39 USD per month, 1,500 premium requests | Opus 4.7 with 7.5x multiplier until 30 April 2026 |
| Copilot Business | 19 USD per user per month, flat request allowance | 19 USD + 30 USD pooled AI credits, token billing |
| Copilot Enterprise | 39 USD per user per month, flat request allowance | 39 USD + 70 USD pooled AI credits, token billing |
| Claude Opus 4.7 (token) | Included in flat fee | 5 USD per million input, 25 USD per million output |
| Claude Code in Pro plan (20 USD) | Included, no hard limits | Test removal for 2 percent of signups, session and weekly caps |
| Refund window Copilot Pro | Not available | Special cancellation with refund until 20 May 2026 |
Notable is the switch from requests to tokens as the billing unit. A request was a flat-rate item, a token is a variable consumption item. A short autocomplete prompt consumes a few hundred tokens, an agent session with multi-file edits and tool calls can run into hundreds of thousands. The bill decouples from interaction count entirely.
Anthropic just shifted enterprise billing from flat-fee to per-token. Every major provider is expected to follow within six months.
Engineering analysis on dev.to, April 2026European perspective: Mid-market, works councils and budgets
European companies are affected on two sides. First, running costs rise. Second, token billing with per-user consumption data creates labour law questions that need to be resolved before rollout. For the 41 percent of German companies with 20+ employees already using AI according to Bitkom, 2026 budgeting becomes more complex.
What this means in practice
- Involve the works council: Per-developer token tracking potentially records performance data. Rollout should be backed by a works agreement.
- GDPR perspective: Token logs can be personally identifiable when linked to user IDs. Retention and purpose belong in the processing register.
- EU AI Act training obligation: Since 2 February 2025 companies must ensure employees working with AI systems have basic AI literacy. The new tariffs make training a budget line.
- Public procurement: IT procurement in public institutions must reflect token tariffs in tenders. Pure per-user flat fees fall short.
Anyone collecting per-developer usage data in Germany can fall under Section 87 paragraph 1 number 6 BetrVG. The works council has co-determination rights. This should be clarified before rollout.
Compared to the US market, European teams have the advantage that works councils and GDPR force structural conversations. This slows rollout but protects against hasty decisions. Starting the dialogue early builds trust and allows clearer rules on token limits, opt-outs and anonymisation.
Why the flat-fee model no longer works
The economic basis of the flat fee is broken for agentic workflows. A classic chat subscriber consumes a few thousand tokens per day, an agent reading a whole codebase and performing multi-file edits runs into hundreds of thousands per session. The equation flat fee divided by average use only works when variance stays small. With agentic use it has exploded.
A Max 20x Claude Code user can consume the equivalent of 600 to 1,500 USD in API tokens for a 200 USD flat fee, according to The Register. Across a customer base that creates an unsustainable loss. Anthropic has responded with weekly caps, Microsoft with credit pools, both are intermediate steps toward full usage billing. OpenAI doubled its GPT-5.5 API pricing in parallel on 23 April 2026, confirming the industry direction.
Any 2026 AI coding tool budget built on flat-fee assumptions has been built on sand. Serious planning from here on needs token metrics per user, per project and per model.
Challenges and risks
The shift comes with concrete risks that go beyond sheer cost increase. Focusing only on the USD number misses at least five knock-on effects that touch teams and strategy. The overview below sorts them by urgency.
Budget uncertainty
Uber's CTO already said his Claude Code budget has been "blown away" according to dev.to. Without token limits, quarterly overruns in the five-figure range become likely.
Works council leverage
Per-user token billing can be classified as performance monitoring. Without a works agreement, rollout can be halted or blocked by interim injunction.
Two-tier teams
Opus 4.7 only in Pro+, junior plans lose access to top models. That hampers learning curves and can cause onboarding disadvantages.
Quality drift
Anthropic admitted on 24 April that Claude Code worked worse for weeks. Under token billing users directly pay for failed agent runs.
Specific risk of vendor lock-in: Token prices are provider-specific and models use different multipliers. Migration becomes harder to cost when workflows are optimised for proprietary features like Claude Code Skills or Copilot Agents. Teams aiming to reduce dependency should evaluate multi-model gateways and deliberately keep contract terms short.
What companies should do now
The next 60 to 90 days are the window to rebuild your AI coding tool budget before token billing becomes the default. The following six steps are ordered by priority and can run largely in parallel. Installing consumption metrics and evaluating alternatives now means acting prepared rather than reactive.
-
Introduce consumption transparency
Measure token use per developer and project before it is billed. GitHub now shows limits in VS Code and Copilot CLI, Anthropic offers similar displays. Add middleware like Portkey or fal.ai as a gateway for unified reporting.
-
Set runtime guards
Budget Guard (hard USD caps per session), Loop Guard (detection of redundant tool calls) and Timeout Guard (wall-clock stop after N minutes) reduce cost measurably. These controls operate at process level, not on monitoring dashboards.
-
Involve the works council early
Per-user consumption data is potentially subject to co-determination. A works agreement for AI tools with clear rules on logging, retention and anonymisation creates legal certainty and trust.
-
Evaluate open-source alternatives
Local models via Continue.dev, Ollama, Qwen or Kimi K2.6 are credible hedges against price spikes. Local setups also deliver GDPR benefits for security-critical codebases. See our analysis of Kimi K2.6 open-weight agents in the enterprise .
-
Role-based allocation
Instead of blanket Pro+ licences, selectively allocate them to senior teams with high agent use. Junior teams can start on Sonnet tariffs and upgrade later. That reduces monthly base costs by 30 to 50 percent.
-
Check exit clauses
GitHub grants refunds for Copilot Pro and Pro+ cancellations until 20 May 2026. Use that window for contract adjustments, switching to Business tariff or evaluating alternatives. For Claude Code, check current contracts for weekly and session caps.
Further reading
Frequently asked questions
Microsoft is moving GitHub Copilot to token-based billing from June 2026. Copilot Business costs 19 USD per user per month including 30 USD in pooled AI credits, Copilot Enterprise costs 39 USD per user per month with 70 USD in credits. The previous flat allowance per request is replaced by actual token consumption per model and session. The change was reported on 23 April 2026 via internal documents seen by the trade blog wheresyoured.at.
On 21 April 2026 Anthropic removed Claude Code from the 20 USD Pro plan on its pricing page and reversed the change the same day after community backlash. Head of Growth Amol Avasare confirmed a test on roughly 2 percent of new prosumer signups. Flat-fee subscriptions miss the actual token cost for heavy users by up to a factor of 10. Anthropic continues to explore ways to keep Claude Code in the low-cost tier without eroding unit economics.
With token billing every agent session costs money based on actual consumption. Agent tasks that read codebases and make multi-file edits run at 20 to 50 USD per day according to morphllm.com, which is 400 to 1,000 USD per month for heavy users. A team of 20 engineers with moderate usage can generate 8,000 to 15,000 USD per month depending on the model and session length. Budget guards, loop detection and timeout controls are required to keep costs in check.
If token usage is tracked per developer, it can fall under Section 87 paragraph 1 number 6 of the German Works Constitution Act (BetrVG) because technical systems that monitor employee behaviour or performance require co-determination. Similar provisions exist in other EU member states via European Works Council directives. Companies should involve the works council before rollout and either anonymise usage data or agree on clear limits to its use. The shift is a good moment to establish a dedicated works agreement for AI tools.
Open-source models like Qwen, Kimi K2.6 and Hermes combined with local inference setups via Continue.dev, Ollama or LM Studio are serious hedges against price spikes. Proprietary alternatives include OpenAI Codex with GPT-5.5 and Google Antigravity. The decisive factor is tool lock-in: teams using multi-model gateways like fal.ai or Portkey stay flexible. For security-critical codebases, local inference also delivers GDPR advantages.
GitHub grants Copilot Pro and Pro+ customers a special cancellation with refund until 20 May 2026 via the Billing settings. The measure was announced on 20 April 2026 and gives users an exit window before limits and model removals tighten for existing subscribers. Using this window preserves the option to switch quickly to Copilot Business or an alternative provider.