AI DevelopmentDecision Matrix9 min readPublished

Gemini 4 Neon vs GPT-6.2 Sol vs Claude Sonnet 5.5 Pricing

The same 100 coding requests can cost less on Gemini than on Sol, even at Gemini’s higher headline price. Prompt length, cache reuse and the agent running the job explain the difference.

DA
Digital Applied
PublishedNovember 19, 2026
Read time9 min

A routine code review with 20,000 input tokens and 2,000 billed output tokens costs six cents on GPT-6.2 Sol or Claude Sonnet 5.5. Google’s Gemini 4 Neon, released November 18, costs nine cents. Now give each model 300,000 input tokens to trace a change across a repository, and allow 10,000 output tokens. The ordering changes. Sonnet costs $0.70, Neon $1.05 and Sol $1.35. All six prices assume ordinary uncached input on the direct standard APIs.

Two questions decide which result matters for your team. How often does a request grow large enough to change the rate? And how many requests does each agent need before a reviewer accepts the change? Token prices answer the first question when paired with your usage log. The second requires knowing which agent, tools and effort setting produced the result.

Key takeaways
  1. 01
    Sol and Sonnet tie on short text requestsBoth charge $2 per million uncached input tokens and $10 per million output tokens. Neon costs 50% more at equal short-request token counts.
  2. 02
    Sonnet has the lowest long-context rates of the threeAbove 272K input tokens, Sol reprices the full request. Neon keeps its $3/$15 rates and Sonnet stays at $2/$10 across its supported window.
  3. 03
    Neon can work directly from a recordingNative audio and video input can eliminate an upstream conversion step. That matters when the timing or sound carries evidence a transcript would lose.
  4. 04
    Compare completed jobs at named effort levelsArtificial Analysis reports different winners for quality, time to first answer and cost per coding task. Track acceptance and review time alongside the invoice.

01 — Rate cardsWhat a million tokens buys

The input column covers material the model receives, including tool results you send back. The output column pays for generated tokens and can include reasoning before the visible answer. The table uses US dollars on the direct APIs as of November 19, 2026. Cache reads, writes and tools have their own prices. Opus and Astra appear as budget references for a harder second attempt.

Uncached standard rates, USD per million input / output tokens. The last column assumes the request fits the model’s window. On OpenAI the threshold is strictly greater than 272K input tokens.
ModelAt or below 272K inputAbove 272K input
GPT-6.2 Sol$2 / $10$4 / $15
Claude Sonnet 5.5$2 / $10$2 / $10
Gemini 4 Neon$3 / $15$3 / $15
Claude Opus 5.5 (step-up)$4 / $20$4 / $20
GPT-6.2 Astra (step-up)$8 / $40$16 / $60

OpenAI applies the long-context input and output rates to the whole request. Anthropic explicitly includes the full 1M context at standard pricing for Sonnet 5.5 and Opus 5.5 under global routing. Google publishes one standard price for Neon. For your own invoice, count input and output separately and apply the rate to each request before adding the costs. OpenAI pricing · Anthropic pricing · Google pricing

02 — Worked examplePrice 100 jobs before averaging

Imagine 90 routine reviews of the size in the opening example. Another ten require the larger repository context. Each job takes exactly one request, with no cache reads or writes. On Sol, explicit cache mode with no breakpoints reproduces that assumption. Keep the token counts fixed across vendors to isolate the effect of the rate card. The 100 jobs contain 4.8 million input tokens, an average of just 48,000 per request. A buyer looking only at that average would put every job in Sol’s short band.

The ten large reviews change the arithmetic. For one of them Sol charges 0.3 × $4 for input and 0.01 × $15 for output, or $1.35. On Neon the output costs the same fifteen cents while the input costs ninety cents. Across ten such jobs, that thirty-cent difference saves $3. Across the other ninety, Neon’s three-cent premium costs $2.70. For this particular queue, Neon finishes thirty cents ahead of Sol.

Claude Sonnet 5.5
Lowest cost at fixed token counts
$12.40

Ninety short requests cost $5.40. The ten larger ones add $7.00.

100 standard requests
Gemini 4 Neon
One price across the queue
$18.60

The short requests cost $8.10 and the longer requests add $10.50.

100 standard requests
GPT-6.2 Sol
The long tail changes the bill
$18.90

The short requests cost $5.40 and the long ones add $13.50.

100 standard requests

Billing all 4.8 million input and 280,000 output tokens at Sol’s short rates would predict $12.40, understating this bill by $6.50. A second mistake runs the other way. If an agent splits one investigation into several turns, classify each turn separately. Carrying 300,000 tokens across a whole job only triggers the premium when an individual request crosses the input threshold. Replaying the same long history on successive turns can trigger the premium several times.

03 — Repeated inputHow much does the second read cost?

A coding agent might send the same contributor guide and tool definitions with every turn. A prompt cache can reuse the computation for that unchanged opening sequence. Neon charges $0.15 per million tokens on a cache hit and uses implicit caching only. Sol and Sonnet both quote $0.10 for a million cache-read tokens on a short request. Sol’s read rate doubles to $0.20 in the long band. An identical advertised discount can therefore produce three different bills.

Consider an unchanged 100,000-token handbook on Sonnet. Reading it cold costs $0.20. A five-minute cache write costs $0.25, followed by $0.01 for a hit. Two uses within that window cost $0.26 versus $0.40 for two uncached inputs. One use costs five cents extra if you write the handbook and never reuse it. A one-hour write costs $0.40 for that same prefix, so even one subsequent hit leaves the two-use total one cent above the uncached total. The third use brings it below. These figures cover the handbook alone.

Sol quotes $2.50 per million for a 30-minute write and $5 for a 12-hour write at short context. Google offers no equivalent explicit write on Neon. For a bill estimate, inspect actual write and hit counts and keep newly appended input in its own column. Check the timing between calls as well. A busy burst of automated reviews and an engineer returning after lunch will reuse different amounts of the same handbook. OpenAI prompt caching · GPT-6.2 migration checklist

04 — EvidenceCan the model read the original?

Picture a bug report recorded on a phone. A customer taps Save twice while a confirmation tone is still playing. A screenshot captures the dialog and a transcript captures what the customer said. The order of the taps, the tone and the dialog disappearing may explain the failure. Neon accepts the original video and audio alongside text, images and PDFs. Sol and Sonnet accept text and images. For either one you can supply selected frames, a timestamped transcript and a log, but somebody or something has to produce those inputs and decide what to keep.

Where the original evidence is plain text, look at length and output limits instead. Neon’s 1,048,576-token budget is shared across input and generation. Extra source files leave less room for its response. Sol lists a 1,050,000-token window with separate maxima of 922,000 input and 128,000 output. Sonnet lists 1M context and up to 128K synchronous output. Its Message Batches API also offers a 300K output beta. If all you need back is a three-line correction, a larger generation allowance has little purchase value.

For current library behavior, supply the relevant documentation or source revision. Google DeepMind gives Neon a March 2026 knowledge cutoff. Sol’s published cutoff is June 1, 2026 and Anthropic lists June 2026 for Sonnet. Sonnet launched on September 28, Sol on November 17 and Neon on November 18. A model can read a release note supplied in context even when its trained knowledge predates that release. DeepMind model card · Sol specifications · Sonnet specifications · Neon launch guide

05 — Interactive useHow long until you see an answer?

Artificial Analysis measured all three models at the setting each vendor calls high. On its November 19 model pages, Neon leads this group on Intelligence Index v5.0.1, a composite across nine evaluations. Sonnet delivers the first answer token soonest. Sol has the shortest reported time per index task. These results can coexist because the models spend different amounts of time reasoning and produce different quantities of output. Each vendor also chooses what its high setting means.

Artificial Analysis model pages, retrieved November 19, 2026. All rows use high effort. Sonnet is the configuration labeled “Default Fallback.” Time per index task is the publisher’s weighted estimate.
Model at highIntelligence IndexFirst answer tokenOutput tokens/secMinutes/task
Gemini 4 Neon6040.64 s10114.8
GPT-6.2 Sol5511.14 s724.7
Claude Sonnet 5.5429.54 s1357.2

For inline help, the wait to read the first useful words matters. For a pull request prepared while a developer is in a meeting, measure when the completed change is ready for review. Artificial Analysis includes reasoning tokens in end-to-end output speed, so dividing the visible patch’s token count by that rate will give a poor estimate of the developer’s wait. Neon results · Sol results · Sonnet results

06 — Repository workWhat changes with a coding agent?

The model inside a coding agent can inspect files, issue commands, read results and revise a patch. The program around it controls which actions are available and what the next turn contains. Artificial Analysis measures the combination in its Coding Agent Index v2.1. Its November 19 tables list Antigravity CLI with Neon at xhigh, Codex with Sol at several efforts, and Claude Code with Sonnet at max with default fallback. There is no equivalent three-way high-effort comparison in those tables.

Antigravity CLI with Neon xhigh records 59 points and $6.86 per task. Claude Code with Sonnet max records 56 and $11.40. Codex with Sol xhigh records 50 and $0.90. Scores are rounded as displayed in the comparison tables. On this task set Sonnet costs more per coding task than Neon despite its lower list prices. Sol xhigh is the least costly of these three configurations, with a nine-point gap to Neon on the composite. The numbers apply to the named tools, efforts and tasks. Antigravity CLI and Codex results · Antigravity CLI and Claude Code results

Cost per accepted change

A $0.90 average per benchmark task tells you what the attempts cost. For a maintenance queue, divide total spend on all attempts by the number of changes a reviewer accepted. Track elapsed time and review minutes beside that number. A cheap patch that takes half an hour to untangle has consumed something the API bill never sees.

07 — SchedulingHow much is the deadline worth?

Suppose the 90 short review requests can run overnight. Their outputs are independent and nobody is waiting for a tool call to return. The published Batch API rates are $1/$5 per million input/output tokens for short Sol requests, $1/$5 for Sonnet and $1.50/$7.50 for Neon. The 90-job portion therefore costs $2.70 on Sol or Sonnet, or $4.05 on Neon, assuming the same token quantities as before. Sol’s long-context Batch rates are a separate $2/$7.50 row.

OpenAI also offers discounted completion windows for eligible background requests. The two-hour target reduces standard model token rates by 30% and the eight-hour target by 40%. On that six-cent short Sol example, the two targets give 4.2 cents and 3.6 cents respectively. The Batch price is three cents. These are distinct ways to submit the job. A background response can support an ongoing workflow, where your application handles a tool call and continues. Price eligible follow-ups as requests in their own right.

Put jobs on the slower route only if the recipient can tolerate the delay. An estimate due before a support call and a pile of pull requests waiting for tomorrow’s stand-up may contain identical input. Giving both jobs the same deadline buys speed for one or misses the deadline for the other. OpenAI background completion windows · GPT-6.2 release and access guide

08 — Buying decisionWhich jobs earn a second attempt?

Pick past tasks for which you know what a good answer looks like. Include a small edit with a clear test, one investigation across several files and one job that previously needed substantial human rework. Run each through the tools your team actually plans to use. Give every candidate the same acceptance criteria, record the effort setting and account for retries. For Sol, check that your integration uses the Responses API for tool calls and requests a supported effort. The model rejects an effort of none.

Reserve an Opus or Astra trial for a problem with a specific observed failure, such as an apparently correct migration that mishandles records written by the old version. Rerun from the original starting point with the same acceptance test. If the premium model succeeds, the value of that result is the avoided repair or the ability to deliver a previously blocked job. If all three cheap models succeed on the first try, raising the budget on the entire queue has little justification. Astra and Sol compared with the premium Claude models · Sonnet and Opus cost per task

Short text or code changes
Compare Sol and Sonnet on first-pass acceptance in the tools your developers already use. Their base token prices tie.
Sol or Sonnet
Long text investigations
Use Sonnet as the cost baseline. Test Neon if better results could offset its premium. Price individual Sol requests that exceed 272K input separately.
Start with Sonnet
Recordings carry evidence
Give Neon the original and compare the accepted result against the full cost of converting the recording and using a text-and-image model.
Try Neon
Decision

Keep the model, the bill and the accepted result together

For a new text-only workflow, Sol and Sonnet give you a six-cent comparison on the example review. For a long text request, Sonnet sets the price to beat. Neon earns a trial where its original media inputs or measured quality can eliminate work. A small record of input size, billed tokens, effort, elapsed time and reviewer acceptance will tell you when those starting choices should change. For help running that comparison against your own queue, see our AI transformation services.

Digital Applied

Compare models on your own review queue

Use representative tickets, API usage and reviewer decisions to price the workflow your team will actually run.

API costsAgent evaluationWorkflow design
Evaluation inputs

Bring a recent change

  • →The original ticket and files
  • →The accepted patch or answer
  • →Usage records and review time
Answers

Frequently asked questions

At or below 272K input tokens on the standard direct APIs, Sol and Sonnet share the lowest uncached rates in this comparison. On longer requests Sonnet retains those rates. The cheapest model for a completed change depends on how many tokens and turns it needs and whether the result passes review.

Digital Applied newsletter

Deep dives on AI, marketing and development.

Practical guides and fresh insights by email. No recycled takes.

Related dispatches

Related reading

AI Development

Is Claude Fable 5.1 Still Worth It for Agentic Coding?

Opus 5.5 outscores Fable 5.1 on every coding benchmark Anthropic published, at 40% of the token price. When Fable 5.1 still earns its place.

September 22, 2026 · 8 minRead
AI Development

Grok 4.7 Costs the Same as 4.6. What Actually Changed

xAI released Grok 4.7 on September 21, 2026 at Grok 4.6's $2/$6 list price. The vendor-run benchmark table, the route price you are billed, three questions.

September 21, 2026 · 7 minRead
AI Development

Gemini 4 Neon vs Hydrogen and Carbon: Which Tier Fits

Compare four Gemini 4 models at November and December API prices. See where Neon fits, when Hydrogen saves money and who gets Orchestrator in Antigravity.

November 19, 2026 · 8 minRead
AI Development

Microsoft MAI Realtime 1: $0.01 Audio Minutes Plus Reasoning

Price Microsoft’s MAI Realtime 1 with its separate reasoning fees, compare voice API billing units and see what teams can test in the limited Foundry preview.

November 19, 2026 · 8 minRead
AI Development

Kimi K2.8 Preview: What Changes for Kimi Code Users?

Check Kimi K2.8 Preview access, context limits and reasoning settings, then plan a coding trial that separates official specifications from performance claims.

September 13, 2026 · 7 minRead
AI Development

Building Games With Astra: A Practical Playtest Guide

Plan an Astra-assisted game around a playable loop, repeatable checks and human feedback. Use a practical playtest method before investing in more polish.

September 9, 2026 · 6 minRead
Google Search

See more Digital Applied analysis in your Google results by adding us as a preferred source.

Add as a preferred source