Gemini 4 Neon vs Claude Opus 5.5
Would switching a repository-editing or terminal agent from Opus 5.5 to Neon improve its results at a lower API rate?
01 // VERDICT
The short answer
| At a glance | Gemini 4 Neon | Claude Opus 5.5 |
|---|---|---|
| List priceinput / output per 1M tokens | $3 / $15 (better) | $4 / $20 |
Highlighted values mark the lower rate or stronger measured score in the same documented run. · Gemini pricing snapshot (verified 2026-11-19)
EXECUTIVE SUMMARY
Try Neon on repository edits if its lower Standard rates matter to your budget. DeepMind’s November 18 table gives Neon the higher CursorBench 4.1 score and Opus a small edge on Terminal-Bench 4.1. The terminal scores came from different sources. An existing Opus deployment is a sensible control for a trial with your own CLI and tests.
Choose Gemini 4 Neon
Your work resembles Cursor-style repository edits and a trial confirms acceptable patches at the lower token rates.
Choose Claude Opus 5.5
Your terminal workflows already pass their acceptance tests, or your integration depends on explicit cache writes with a chosen lifetime.
02 // KEY FACTS
Key facts
- Cursor supplied both CursorBench 4.1 scores to DeepMind. The model card lists 65.4% for Neon and 58.1% for Opus 5.5. The published table does not establish identical run settings.
- On Terminal-Bench 4.1 the cheating-adjusted scores were 65.2% for Neon and 65.8% for Opus. Artificial Analysis ran Neon with Antigravity-CLI 1.5.9. DeepMind took the Opus figure from the public leaderboard.
- Standard rates on November 19, 2026 were $3 fresh input, $0.15 cached input and $15 output per million tokens for Neon. Anthropic listed $4, $0.20 and $20 for Opus. Opus also offers explicit cache writes at $5 for five minutes or $8 for one hour per million tokens.
- For 10,000 requests with 12,000 fresh input, 8,000 cached input, zero writes and 4,000 billed output tokens each, the Standard token estimate is $972 on Neon and $1,296 on Opus. Include retries and cache creation when comparing spend per accepted patch.
03 // TIER PRICING
API rates by service tier
USD per 1M tokens from vendor pricing pages. 'Not published' means exactly that — we don't extrapolate.
| Service tier | Gemini 4 Neon | Claude Opus 5.5 |
|---|---|---|
| Standarddefault on-demand serving | in $3 · cached $0.15 · out $15 implicit cache only (no write fee) | in $4 · cached $0.20 · out $20 cache write $5 |
| Batchasync jobs, 50% off | in $1.5 · cached $0.075 · out $7.5 implicit cache only (no write fee) | in $2 · cached $0.10 · out $10 cache write $2.5 · derived |
| Flexslower/cheaper serving, 50% off | in $1.5 · cached $0.075 · out $7.5 implicit cache only (no write fee) | Not published |
| Fast2× on supported OpenAI and Anthropic models | Not published | in $8 · cached $0.40 · out $40 cache write $10 · derived |
| Long-context band | Gemini 4 Neon | Claude Opus 5.5 |
|---|---|---|
| Long-context ratesapplies per request above the model threshold | Flat pricing same rates at any input size (up to context window) | Flat pricing same rates at any input size (up to context window) |
USD per 1M tokens. Every input token is billed in exactly one bucket — uncached, cached read, or cache write; a write replaces the input rate for those tokens (official rule), so write rates are per tier as listed. Anthropic rows bill the 5-minute cache-write TTL ($5/MTok on Opus 5.5, $10/MTok on Fable 5.2); 1-hour writes ($8 and $16) are also published, and cache multipliers stack with Batch and Fast-mode pricing. · Anthropic pricing (verified 2026-11-19) · These Gemini models use implicit caching. Enter cache hits as cached input and set explicit cache writes to zero. Gemini 4 Neon has a fixed published rate for each tier. · Google Gemini pricing (verified 2026-11-19)
Gemini service tiers
04 // COST SIMULATOR
Run this comparison on your own numbers
Preloaded with a workload typical for this comparison. Change anything — results recompute locally.
Estimated monthly ledger
browser-localToken-based workload estimate only; it does not measure quality or cost per successful task.
| Model | Per request | Monthly |
|---|---|---|
| Gemini 4 Neoncheapest | $0.0972 | $972.00 |
| Claude Opus 5.5 | $0.1296 | $1,296.00 |
Monthly spread between the cheapest and priciest selected model: $324.00.
Breakdown: Gemini 4 Neon · standard tier · standard context
- Fresh input (uncached): 12,000 tok × $3/M$0.036
- Cached input (reads): 8,000 tok × $0.15/M$0.0012
- Cache writes: 0 tok · rate unavailable$0.00
- Output (incl. reasoning tokens): 4,000 tok × $15/M$0.06
- Per request$0.0972
Rates: official pricing · verified 2026-11-19
- Per-token arithmetic only: it does not model quality differences or cross-vendor output-token-count differences for the same task.
- Reasoning tokens are billed as output tokens at the chosen model's output rate.
- Long-context band is decided per request by total input tokens vs the provider threshold, not by monthly volume.
- Cache accounting uses the official mutually-exclusive buckets: each input token is billed once, as uncached input, cached read, OR cache write — writes replace the input rate for those tokens, never add to it.
- Claude rows bill the 5-minute-TTL cache write; the published 1-hour TTL ($8/MTok standard) is not modeled (no TTL input).
- Rows flagged 'derived' are computed from a stated official rule (e.g. Anthropic batch -50%), not read verbatim from a price table.
- Model and tier combinations without a published rate appear as Unsupported.
05 // BENCHMARK LOG
Benchmark results and conditions
- CursorBench 4.1Gemini 4 Neon
- 65.4%
Cursor supplied the score to Google DeepMind for its November 18 model card.
- Terminal-Bench 4.1Gemini 4 Neon
- 65.2%
Artificial Analysis ran Neon with Antigravity-CLI 1.5.9. DeepMind reports cheating-adjusted accuracy. DeepMind compiled the scores from different sources. A shared run is unconfirmed.
- CursorBench 4.1Claude Opus 5.5
- 58.1%
Cursor supplied the score to Google DeepMind for its November 18 model card.
- Terminal-Bench 4.1Claude Opus 5.5
- 65.8%
DeepMind took Opus’s cheating-adjusted accuracy from the public Terminal-Bench leaderboard. DeepMind compiled the scores from different sources. A shared run is unconfirmed.
06 // BY WORKLOAD
Which model for which workload
Confidence labels: Measured = same-harness evidence · Price-derived = list-price arithmetic · Inferred = reasoned from docs · Insufficient evidence = no comparable public data.
| Workload | Our pick | Confidence | Why |
|---|---|---|---|
| Daily coding tasks | Gemini 4 Neon | Inferred | Cursor’s reported score makes Neon a candidate for an editing trial. Require the same patch review and tests before counting a task as solved. |
| Complex repository maintenance | Decide by bake-off | Insufficient evidence | A 0.6-point separation in compiled terminal results cannot predict the result of switching the model inside your CLI. Reuse the same tool permissions, time limit and tests for both. |
| Budget-sensitive / high volume | Gemini 4 Neon | Price-derived | The November 19 fresh-input, cached-read and output rates are each 25% lower on Neon. Matching the token mix gives the same 25% reduction in the worked budget. |
QUESTIONS
Frequently asked questions
Which model scored higher on CursorBench and Terminal-Bench?
In DeepMind’s November 2026 table Neon led CursorBench 4.1, 65.4% to 58.1%. Opus led Terminal-Bench 4.1, 65.8% to 65.2%. The evaluations cover different coding tasks and use different reporting paths.
Did DeepMind run both models in Antigravity-CLI for Terminal-Bench?
Its methodology credits Artificial Analysis with running Neon on Antigravity-CLI 1.5.9. The Opus score comes from the public leaderboard. Cursor supplied the separate CursorBench scores. A matched CLI run would be a separate test.
How should I compare cache-heavy code reviews?
Use each provider’s actual cache-hit counts. Neon uses implicit cache hits. For Opus, include explicit writes at the TTL your application requests. Record billed output and failed attempts alongside the cost of each accepted patch.
CONTINUE EXPLORING
Related
VERIFICATION LOG
Sources & freshness
- Googleaccessed 2026-11-19Gemini 4 Neon API model documentationhttps://aistudio.google.com/docs/models/gemini-4-neonModel ID, token limits, implicit caching, thinking levels and supported tools.
- Anthropicaccessed 2026-10-09Claude models overview (model IDs, context window, max output, retirement)https://platform.claude.com/docs/en/docs/about-claude/models/overviewclaude-opus-5-5: 1M context, 128K max output, training cutoff Jun 2026, retirement not sooner than 2027-09-22, cache reads '5% of input price', batch 50% off.
- Google DeepMindaccessed 2026-11-19Gemini 4 Neon model cardhttps://deepmind.google/models/model-cards/gemini-4-neon/ · published 2026-11-18DeepMind identifies Carbon as the basis for Neon and dates the knowledge cutoff to March 2026. The evaluation table includes CursorBench 4.1 and cheating-adjusted Terminal-Bench 4.1.
- Google DeepMindaccessed 2026-11-19Gemini 4 Neon model evaluation (methodology and results)https://storage.googleapis.com/deepmind-media/gemini/gemini_4_neon_model_evaluation.pdfThe results table is on the page numbered 5. The agentic coding methodology credits Cursor for CursorBench and distinguishes Neon's Artificial Analysis terminal run from the other public leaderboard results.
- Anthropicaccessed 2026-11-19Claude pricing docs (cache TTL prices, fast/batch stacking rules)https://platform.claude.com/docs/en/about-claude/pricingOpus 5.5 1h cache write $8/MTok; caching multipliers apply on top of fast-mode pricing and stack with the Batch discount; fast mode not available with Batch. Re-verified 2026-11-18: Fable 5.2 Batch 50% off ($4/$20), Fast 2x ($16/$80), 2.5% cache-read stacking, 1h cache write $16.00.
- Anthropicaccessed 2026-11-19Claude pricing (API pricing table incl. cache write/read, fast mode, batch)https://claude.com/pricingOpus 5.5: $4/$20 per MTok, cache read $0.20, cache write (5m TTL) $5, batch 'save 50%', fast mode 2x on input/output. Re-verified 2026-11-18: Claude Fable 5.2 ($8/$40, $0.20 cached read = 2.5%, $10.00 5m write).
- Googleaccessed 2026-11-19Gemini Developer API pricinghttps://aistudio.google.com/docs/pricingThe paid tables distinguish fresh input, implicit cache hits and output including thinking. Argon and Carbon have introductory Standard, Batch and Flex rates through November 30, 2026, with the next prices effective December 1. Google lists Priority separately.
Pricing data last verified 2026-11-19. Vendors ship frequently — re-verify before making spend commitments. See the methodology page for update cadence and stale-data flags.