Measured 2026-04-23 · 15 production models

Up to 38% off your flagship LLM bill.

Measured on 2026-04-23: Opus 4.7, GPT-5.4, and Gemini 3.1 Pro.

We sent the same prompt to each vendor's flagship, compressed and uncompressed, and recorded the input_tokens they actually billed. Opus 4.7 billed 38.3% less · GPT-5.4 billed 35.3% less · Gemini 3.1 Pro billed 34.6% less. Same prompt, live API calls, on 2026-04-23.

Get free API key Integration docs

1,000 compressions/mo free. No card required.

Billed token counts2026-04-23

Claude Opus 4.7−38.3%

992 → 612 input_tokens

GPT-5.4−35.3%

515 → 333 prompt_tokens

Gemini 3.1 Pro−34.6%

566 → 370 promptTokenCount

See the full methodology →

How it works

Same answer. Shorter prompt. Smaller bill.

Before your prompt reaches Claude, GPT, or Gemini, gotcontext spots repeated ideas, filler phrases, and redundant context, then rewrites the prompt to say the same thing with fewer words. The AI gets a shorter prompt, gives the same answer, and your input bill drops ~35%.

§01 / Input

You send a prompt

Same prompt you'd send today. No code changes. gotcontext sits between your app and the LLM. One line to add us, one line to remove us.

§02 / Compress

We compress it

We keep what matters (names, numbers, constraints, instructions) and drop what doesn't: repeated boilerplate, filler words, restated context. Takes about 80ms.

§03 / Output

Your LLM answers normally

The AI reads a shorter version of your prompt and gives the same answer. You pay for fewer input tokens. Across flagships that's 35-38% off your bill.

Before · what you'd send992 tokens

"The FastAPI app initializes Sentry. The FastAPI app wires Clerk auth. The FastAPI app exposes an MCP gateway at /mcp. Circuit breakers guard Redis, Postgres, and Polar. When a breaker opens, DegradationHeader adds X-Degraded. The MCP gateway is at /mcp and validates gc_ API keys..."

After · what gotcontext sends612 tokens (-38%)

"§1=The FastAPI app initializes Sentry, wires Clerk auth, exposes MCP gateway at /mcp. Circuit breakers (Redis/Postgres/Polar) trigger X-Degraded header on open. gc_ API keys validate via verify_api_key..."

§04 / Fidelity

"Does the AI give worse answers?"

No. Every compression is scored for semantic fidelity against the original. If a compression would meaningfully change the answer, we ship the original through untouched. You can dial the fidelity-vs-savings balance per request (fidelity: "minimal" | "balanced" | "full"). Default is balanced, which is what produced the 38% Opus 4.7 number above.

Lossless mode: maximum fidelityBalanced: high fidelity, ~35% savingsAggressive: for high-volume routing / log summaries

Real provider-billed savings

What each flagship actually charged us

We sent the same 279-word prompt to each vendor's flagship -- Opus 4.7, GPT-5.4, and Gemini 3.1 Pro -- once uncompressed, once compressed through gotcontext, and read back the input_tokens field each provider reports. These numbers are what hits your invoice.

Provider	Model tested	Uncompressed	Compressed	Billed savings
Anthropic	claude-opus-4-7	992 tokens	612 tokens	-38.3%
OpenAI	gpt-5.4	515 tokens	333 tokens	-35.3%
Google	gemini-3.1-pro-preview	566 tokens	370 tokens	-34.6%

Three different tokenizers, three different raw counts, same compressed input. Opus 4.7 surfaces the most savings because its tokenizer is the most granular -- it picks up small wins other vendors round off. Our semantic tokenizer reported 39.8% on this corpus. The reproducible benchmark script is open-source. For a head-to-head against the most active OSS competitor, see the gotcontext vs Headroom benchmark (76% to 41% on real OpenAI bills, measured 2026-04-25).

Projected dollar savings across 15 models

Dollar savings per compression: all 15 models

We project the dollar value of one compression per model by applying each provider family's measured billed savings % (Anthropic 38.3%, OpenAI 35.3%, Google 34.6%) to a 691-token reference prompt -- the mean of the three uncompressed counts above. Bars are colored by provider.

Claude Fable 5$0.002647 / call$2,647.00 / month @ 1M

Claude Opus 4.8$0.001323 / call$1,323.00 / month @ 1M

Claude Opus 4.7$0.001323 / call$1,323.00 / month @ 1M

Claude Opus 4.6$0.001323 / call$1,323.00 / month @ 1M

GPT-5.5$0.001220 / call$1,220.00 / month @ 1M

Claude Sonnet 4.6$0.000794 / call$794.00 / month @ 1M

GPT-5.4$0.000610 / call$610.00 / month @ 1M

GPT-4.1$0.000488 / call$488.00 / month @ 1M

Gemini 3.1 Pro$0.000478 / call$478.00 / month @ 1M

Gemini 3.5 Flash$0.000359 / call$359.00 / month @ 1M

Claude Haiku 4.5$0.000265 / call$265.00 / month @ 1M

GPT-5.4 mini$0.000183 / call$183.00 / month @ 1M

Gemini Auto$0.000143 / call$143.00 / month @ 1M

GPT-4.1 mini$0.000098 / call$98.00 / month @ 1M

Gemini 2.5 Flash$0.000072 / call$72.00 / month @ 1M

Anthropic · Claude family

Claude: Opus, Sonnet, Haiku

Claude uses a proprietary tokenizer with 200K or 1M context depending on tier. Compression savings are largest here because Anthropic's premium tier is the most expensive per-token bill of any provider.

Claude Fable 5

ProjectedAnthropic

Mythos-class flagship (2026-06-09). 2x Opus pricing — the strongest compression ROI in the Anthropic lineup.

Input: $10.00/1M
Output: $50.00/1M
Context: 1M
Saved / call: $0.002647

1M calls / mo: $2,647.00

10M calls / mo: $26,470.00

Claude Opus 4.8

Anthropic

Current Anthropic flagship (2026-05-27). Same pricing as 4.7, 1M context.

Input: $5.00/1M
Output: $25.00/1M
Context: 1M
Saved / call: $0.001323

1M calls / mo: $1,323.00

10M calls / mo: $13,230.00

Full breakdown →

Claude Opus 4.7

FlagshipAnthropic

Flagship reasoning. 1M context, predecessor to 4.8.

Input: $5.00/1M
Output: $25.00/1M
Context: 1M
Saved / call: $0.001323

1M calls / mo: $1,323.00

10M calls / mo: $13,230.00

Full breakdown →

Claude Opus 4.6

Anthropic

Premium tier. Pinned predecessor to 4.7.

Input: $5.00/1M
Output: $25.00/1M
Context: 200K
Saved / call: $0.001323

1M calls / mo: $1,323.00

10M calls / mo: $13,230.00

Claude Sonnet 4.6

Anthropic

Balanced cost and quality. The workhorse tier.

Input: $3.00/1M
Output: $15.00/1M
Context: 200K
Saved / call: $0.000794

1M calls / mo: $794.00

10M calls / mo: $7,940.00

Full breakdown →

Claude Haiku 4.5

Anthropic

Low-cost routing and summary tasks.

Input: $1.00/1M
Output: $5.00/1M
Context: 200K
Saved / call: $0.000265

1M calls / mo: $265.00

10M calls / mo: $2,650.00

Full breakdown →

OpenAI · GPT family

GPT: 5.4, 5.4 mini, 4.1, 4.1 mini

OpenAI prices vary 40× between the mini and flagship tiers. GPT-5.4 has a 400K context, useful for very long prompts, and compression amplifies that advantage by letting you fit more source material.

GPT-5.5

FlagshipProjectedOpenAI

Flagship OpenAI model (2026-04-23). Output rate doubled vs GPT-5.4 — compression ROI is highest here.

Input: $5.00/1M
Output: $30.00/1M
Context: 1M
Saved / call: $0.001220

1M calls / mo: $1,220.00

10M calls / mo: $12,200.00

GPT-5.4

OpenAI

Standard OpenAI tier — the model we measured directly.

Input: $2.50/1M
Output: $15.00/1M
Context: 400K
Saved / call: $0.000610

1M calls / mo: $610.00

10M calls / mo: $6,100.00

Full breakdown →

GPT-5.4 mini

OpenAI

Low-cost GPT-5 tier with the same 400K window.

Input: $0.75/1M
Output: $4.50/1M
Context: 400K
Saved / call: $0.000183

1M calls / mo: $183.00

10M calls / mo: $1,830.00

GPT-4.1

OpenAI

Mid-cost OpenAI profile suited to compact prompts.

Input: $2.00/1M
Output: $8.00/1M
Context: 128K
Saved / call: $0.000488

1M calls / mo: $488.00

10M calls / mo: $4,880.00

GPT-4.1 mini

OpenAI

Bulk preprocessing and light classification.

Input: $0.40/1M
Output: $1.60/1M
Context: 128K
Saved / call: $0.000098

1M calls / mo: $98.00

10M calls / mo: $980.00

Google · Gemini family

Gemini: 3.1 Pro, Auto, 3.1 Flash

Gemini ships the cheapest mainstream tier (Flash) and the largest context window (1M across all three). Auto routes between Pro and Flash per turn for a middle-tier effective rate.

Gemini 3.5 Flash

Google

Gemini current Flash flagship (2026-05-19). Flat 1M context at $1.50/$9 per MTok.

Input: $1.50/1M
Output: $9.00/1M
Context: 1M
Saved / call: $0.000359

1M calls / mo: $359.00

10M calls / mo: $3,590.00

Full breakdown →

Gemini 3.1 Pro

FlagshipGoogle

Gemini large-context flagship ($2/$12 <=200K).

Input: $2.00/1M
Output: $12.00/1M
Context: 1M
Saved / call: $0.000478

1M calls / mo: $478.00

10M calls / mo: $4,780.00

Full breakdown →

Gemini Auto

Google

Auto-routes between Pro and Flash per turn.

Input: $0.60/1M
Output: $5.00/1M
Context: 1M
Saved / call: $0.000143

1M calls / mo: $143.00

10M calls / mo: $1,430.00

Gemini 2.5 Flash

Google

Low-cost Gemini tier. Ideal for bulk compression.

Input: $0.30/1M
Output: $2.50/1M
Context: 1M
Saved / call: $0.000072

1M calls / mo: $72.00

10M calls / mo: $720.00

Full breakdown →

Methodology

How we measured

We drove compression through compress_meta_tokens on our production MCP gateway at https://api.gotcontext.ai/mcp. Each call sets _meta.model to the target so the gateway's per-model attribution logs correctly.

Corpus: 279-word technical prose with realistic repetition (FastAPI + MCP documentation). Same prompt sent to all three providers, compressed and uncompressed.
Real-provider calls: Direct API calls to api.anthropic.com/v1/messages, api.openai.com/v1/chat/completions, and generativelanguage.googleapis.com. We read the usage.input_tokens field each response returns, the same number the provider bills.
Semantic-layer measurement: gotcontext's own tokenizer: 279 words compress to 168 tokens (39.8% reduction). We quote this separately because it's what the compressor achieves at the structural level, independent of provider billing.
Pricing source: Last verified 2026-04-23 against each provider's official pricing page. Internal source of truth is token-saver-5000/src/provider_profiles.py; the /v1/models endpoint exposes the same data.

Honest disclaimer: 39.8% is the ratio for this specific corpus. Short documents (<100 tokens) may expand due to skeleton overhead. Medium-to-large documents (500+ tokens) typically compress 5 to 10×. Highly repetitive content (logs, API responses, boilerplate) can hit 85%+. Run your own benchmark on a representative sample before scaling.

Reproduce it

Run the benchmark yourself

Two public scripts. Both are open-source in the main repo. No sign-up needed to reproduce our numbers.

1. Real-provider billing (the headline data)

Hits Claude, GPT, and Gemini with the same compressed + uncompressed prompt. Reads the input_tokens field from each response. Requires your Anthropic / OpenAI / Google keys.

ANTHROPIC_API_KEY=... OPENAI_API_KEY=... GEMINI_API_KEY=... \
python benchmarks/real_llm_cross_provider_smoke.py

benchmarks/real_llm_cross_provider_smoke.py →

2. Per-model attribution (gotcontext side)

Drives compress_meta_tokens once per registered model with _meta.model set. Cross-checks against /v1/usage/by-model.

GC_API_KEY=gc_your_key python benchmarks/per_model_savings_smoke.py

benchmarks/per_model_savings_smoke.py →

Ready to cut your LLM bill?

The difference between Opus 4.7 and GPT-5.4 mini is 37×. Pick the right model, add compression, and stop paying for tokens you never needed.

Start free See pricing