The Opus 4.7 multiplier in GitHub Copilot has tripled in 60 days. GitHub’s changelog1 shows it launched at 7.5× on April 16, jumped to 15× on May 1 when promotional pricing expired, and for annual-plan subscribers staying on request-based billing, reached 27× on June 1. The same day, GitHub moved monthly subscribers to token-metered AI Credits2, billed at published per-token rates. The cost direction is upward under both systems; the per-turn number depends on which plan you’re on.
The 7.5× Launch and the Hidden Cliff
GitHub’s April 16 changelog1 announced Opus 4.7 at a 7.5× premium-request multiplier with promotional pricing “through April 30.” The expiry date was in the announcement; the post-April-30 rate was not. Most subscribers absorbed 7.5× as the stable price.
Copilot’s premium-request system is GitHub’s own accounting layer sitting on top of the Anthropic API. It is not the same as Anthropic’s per-token pricing, which remains unchanged. Per GitHub’s documentation on premium requests3, each model carries a fixed multiplier applied against a plan’s monthly allowance: Sonnet 4.6 at 1×, Opus 4.7 at 7.5× (later 15×, then 27×), and Opus 4.6 in fast mode at 30×. The fast-mode figure shows the multiplier can scale well beyond what Anthropic’s underlying token pricing would suggest.
For a Pro+ plan with 1,500 monthly premium requests, the arithmetic across the promotional and current rates:
| Period | Multiplier | Turns from 1,500 allowance |
|---|---|---|
| Opus 4.6 (removed Apr 20) | 3× | ~500 |
| Opus 4.7 promotional (Apr 16–30) | 7.5× | ~200 |
| Opus 4.7 standard (May 1 onward) | 15× | ~100 |
| Opus 4.7 annual legacy (Jun 1 onward) | 27× | ~55 |
May 1: The 15× Reveal
When promotional pricing ended, the same changelog page updated to 15×. That’s 2× the launch price. Subscribers who had built agentic workflows at 7.5× were now running those loops at twice the PRU cost. Community discussion #1944704 began filling with users re-evaluating Opus 4.7’s place in their tool rotation. One enterprise customer in that thread reported cancelling 30 Copilot Enterprise seats and migrating to Claude Teams.
The disclosure pattern is worth noting. GitHub did publish the 15× figure; it was not leaked. But a changelog editor’s note is not how most engineering teams expect to learn that a tool they scripted into CI pipelines just doubled in price. Community discussion #1928145 documents users doing their own cost math against direct API pricing differences, estimating that an equivalent API cost differential would land around 4×, not 7.5×.
June 1: Annual Plans Hit 27×
The escalation does not stop at 15×. For annual Copilot Pro and Pro+ subscribers staying on request-based billing, GitHub’s multiplier reference6 set Opus 4.7 at 27× and Sonnet 4.6 at 9× effective June 1. The same day, GitHub moved monthly subscribers to token-metered AI Credits2, billed at published per-token rates.
AI Credits vs. PRUs: The Real Cost of an Opus Turn
GitHub’s pricing documentation7 lists these token rates:
| Model | Input (per M tokens) | Output (per M tokens) | PRU multiplier (annual, June 1) |
|---|---|---|---|
| Claude Fable 5 | $10.00 | $50.00 | launched Jun 9, suspended Jun 12 [Updated June 2026] |
| Opus 4.8 | $5.00 | $25.00 | 27× |
| Opus 4.7 | $5.00 | $25.00 | 27× |
| Sonnet 4.6 | $3.00 | $15.00 | 9× |
Claude Fable 5, launched June 9, 2026 as the first Mythos-class model above the Opus tier, prices at $10/$50 per million tokens at the Anthropic API level, exactly 2× Opus 4.8’s $5/$25. GitHub added Fable 5 to Copilot on June 9 — billed at provider list pricing under usage-based billing, with no separate PRU multiplier published — but suspended access across all Copilot experiences on June 12 following an Anthropic announcement.14 The suspension makes any future Copilot multiplier for Fable 5 speculative, but the per-token cost structure (2× Opus 4.8) sets the floor for what any such multiplier would express. [Updated June 2026]
Under token billing, Opus 4.7 costs 1.67× more than Sonnet 4.6 on both input and output. Under PRU billing, the ratio is 3× (27 vs. 9). The two billing systems price the gap between models differently, which matters when choosing between them.
For agentic loops, costs scale linearly with token counts. At GitHub’s published per-token rates7, Opus 4.7 costs 1.67× more than Sonnet 4.6 for the same input and output volume. Community discussion #1929488 places a single heavy autonomous Opus session at $30–40 in API-equivalent cost. On a Pro+ plan’s 7,000 AI Credits ($70 of included usage at $0.01 each, after GitHub’s May 12 flex allotment lifted Pro+ from $39 to $70), a subscriber can realistically run only one or two sustained autonomous sessions before exhausting the monthly allocation. (GitHub Blog. “Flex allotments in Pro and Pro+, and a new Max plan.” May 12, 2026)
Billing Disputes and the Observability Gap
The more operationally disruptive report is in community discussion #1929119, where a user posted billing records showing 7.5× charges attributed to Claude Sonnet 4.6, Claude Haiku 4.5, and Gemini 3 Flash, three models the user states they did not invoke. The charges resulted in roughly an $18 overcharge and exhausted the user’s premium quota ahead of schedule.
This is a single-user report, not an audited billing discrepancy or independently confirmed pattern. But the specifics are concrete: named models, a cited multiplier, documented quota exhaustion. GitHub’s response was to close the discussion and direct the user to open a support ticket, stating the community forum is not a support channel.
The structural issue is not whether the dollar figure is correct; it is what the incident reveals about observability. A developer running agentic workflows cannot easily audit which model each sub-call resolved to, or what multiplier was applied per invocation. If a background agent resolves to a more expensive model than the user explicitly selected, the cost surface is larger than it appears from the session log.
Model Access Is Shrinking: Where Opus Still Lives
GitHub’s April 20 plan changes10 removed Opus models from Copilot Pro entirely and paused new self-serve signups for Pro and Pro+. Opus 4.7 is now a Pro+-only model, and older Opus variants (4.5 and 4.6) are being removed from Pro+ as well. The annual multiplier table6 now lists Claude Opus 4.8, generally available in Copilot since May 28, at 27×, the same rate as 4.7. (GitHub Blog. “Claude Opus 4.8 is generally available for GitHub Copilot.” May 28, 2026) The top-tier Opus slot is still being refreshed, but access to it is narrowing to Pro+ subscribers alone.
Anthropic launched Claude Fable 5 on June 9, 2026 as the first model in the Mythos class, positioned above the Opus tier entirely.13 GitHub added it to Copilot the same day for Pro+, Max, Business, and Enterprise users.14 Three days later, on June 12, GitHub suspended Fable 5 access across all Copilot experiences following an Anthropic announcement — the suspension affected all users regardless of plan.14 [Updated June 2026] All other Claude models including Opus 4.8 and Sonnet 4.6 remained available and unaffected. At $10/$50 per million tokens at the Anthropic API level, Fable 5 carries exactly double Opus 4.8’s underlying cost, which sets the ceiling for any future Copilot multiplier calculation should GitHub reinstate it. The 27× figure that currently governs Opus 4.8 in Copilot was set when Opus 4.8 was the most capable widely available Claude model; it is now the second tier. The cost escalation pattern traced in this article may continue into a fourth step if and when Fable 5 access is restored and GitHub publishes a multiplier. For teams already at the ceiling of what 27× Opus 4.8 can justify, the fast-mode pricing tradeoff in Claude Code and the broader Copilot vs. Cursor vs. Claude Code comparison remain the practical decision anchors.
The Fable 5 Three-Day Window: A Case Study in Model-Access Risk
The Fable 5 episode compresses the Opus 4.7 multiplier story into a sharper form. GitHub added the model on June 9 and removed it on June 12 — a 72-hour window. Teams that built integrations during that window, or that configured routing logic to prefer Fable 5, had those integrations broken with no advance notice. The suspension reason was an external directive to Anthropic, not a GitHub product decision, but the operational impact on Copilot users was the same: a model that existed in the catalog on Monday did not exist on Thursday.
This pattern — rapid capability introduction, rapid removal, varying disclosure timelines — argues for a particular architectural stance: any Copilot workflow that depends on a specific model by name should treat that dependency as fragile. GitHub’s auto model selection (which carries a 10% credit discount) provides platform-managed fallback but removes per-model cost control. The alternative is explicit model pinning with fallback logic owned at the application layer — the same conclusion the multiplier escalation from 7.5× to 27× points toward, for different reasons. For teams weighing how Fable 5’s credit billing compares across the direct API and subscription tiers, the billing mechanics for the June 23 credit shift cover the per-token cost implications in detail.
Three Escape Routes for Agentic Teams
Teams that integrated Opus 4.7 into autonomous loops face a forced choice.
Take the AI Credits route. For monthly plans this became the default on June 1; it eliminates the multiplier algebra and bills at Anthropic’s API rate. Credits are finite and non-refundable; a runaway agent or recursive tool call can zero out a monthly balance in one afternoon, with no fallback to cheaper models once credits are gone.
Fall back to Sonnet 4.6. Cheaper at 1.67× the per-token rate of Opus (token billing) or 3× (PRU billing, 9× vs. 27×). Different capability tier. Teams that selected Opus specifically for long-context reasoning or complex multi-file refactors will notice the downgrade.
Rewrite routing logic. Gate Opus behind human approval, cap iteration counts, or shard work across cheaper models. A hybrid approach limits exposure: route Sonnet 4.6 to tool calls and context accumulation, reserve Opus 4.7 for the turns requiring deep inference. The slow route, but the only one that preserves the original capability without accepting the new pricing in full.
The Broader Lesson: Launch Promos Are Not Budgets
The 7.5× launch rate was always labeled as promotional. The problem is that “promotional” in vendor copy reads as “introductory” to buyers, who treat it as a stable baseline for headcount planning and CI budgets. When the actual rate turned out to be double, then triple, then nearly quadruple the launch price, the shock was not that GitHub raised prices, it was that the floor turned out to be fiction.
Community discussion #1944704 shows the expected distribution: subscription cancellations, evaluation of Claude Code, Cursor, and direct Anthropic API access. The May 1 15× jump triggered most of the activity; the June 1 27× annual-plan escalation has since set off another wave from subscribers who waited.
This is a recurring pattern in AI tooling. Vendors launch capabilities at unsustainable margins to drive adoption, then correct pricing once the integration work is sunk. The second-order consequence for engineering teams is structural: any multiplier, token price, or “credits included” figure announced during a launch window should be treated as marketing copy, not as infrastructure cost to budget against. Build cost models around the post-promotion rate, or build architecture so models can be swapped in an afternoon. Betting that a 7.5× multiplier will hold is betting against the vendor’s incentive to stop subsidizing your compute.
Frequently Asked Questions
What does an Opus 4.7 turn actually cost in dollars under AI Credits?
One AI Credit equals $0.01 USD. At Opus 4.7’s published rates ($5.00/M input, $25.00/M output), a moderate agentic turn of roughly 5,000 input and 2,000 output tokens costs about $0.075. Over 200 such turns in a day, that’s approximately $15 before any caching or context-reuse reductions.
Why is PRU billing worse than token billing for short Opus completions?
The 27× PRU multiplier applies per request regardless of token count. A one-line autocomplete and a full multi-file agent refactor both consume 27 PRUs. Under token billing, the autocomplete costs a fraction of a cent while the refactor costs dollars, so PRU billing makes short turns disproportionately expensive relative to their actual token consumption.
Does Claude Code bypass the Copilot multiplier system entirely?
Yes. Claude Code authenticates directly against Anthropic’s API using your own key, billing at published per-token rates with no Copilot multiplier or credit conversion. The tradeoff is losing Copilot’s inline IDE suggestions, PR-aware chat context, and the fixed PRU allowance bundled into a Pro+ subscription.
Could GitHub raise the multiplier again after June 1?
Nothing in GitHub’s published documentation commits to the 27× rate as a ceiling. The multiplier escalated from 7.5× to 15× to 27× in 60 days with no advance notice of each step, so annual-plan subscribers should treat the current rate as mutable rather than a stable pricing floor.
What happens to an unattended CI agent when the AI Credit balance hits zero mid-session?
Under AI Credits, hitting the cap is a hard stop: the model returns errors and the session terminates, with no automatic degradation to a cheaper model like Sonnet. This is a structural change from request-based billing, where all models draw from the same pooled allowance and a 9× Sonnet request still completes. Teams running agents in CI or deployment pipelines should implement credit-balance pre-checks before launching Opus sessions, or accept that a runaway tool-calling loop can silently kill a pipeline mid-run.
