01
Your browser
The original prompt and identifiers stay on your device.
Original text stays here
Compress the tokens
The only
lossless token compression.
It's also the fastest.
Direct API or GPTZIP Proxy Works with most LLM harnesses Private Mode masking
Before
182
tokens
After
129
tokens
kept removed
182 before129 after
−29.1%
input reduction
27 ms
Compression Time
Counting GPTZIP token savings…
Benchmark data
LLM accuracy increases after GPTZIP compression.
Accuracy and compression
-40pp
-30pp
-20pp
-10pp
0ppNo accuracy drop
0%
10%
20%
30%
40%
50%
60%
70%
Median tokens removedAccuracy change vs uncompressed
GPTZIPHeadroomBear (TheTokenCompany)LLM-Lingua (MSFT)| Benchmark | Method | Median tokens removed | Accuracy change |
|---|---|---|---|
| CoQA | GPTZIP | 13.9% | 0.0pp |
| CoQA | Headroom | 51.8% | −21.6pp |
| CoQA | Bear (TheTokenCompany) | 13.1% | −0.1pp |
| CoQA | LLM-Lingua (MSFT) | 63.0% | −18.5pp |
| SQuAD v2 | GPTZIP | 12.3% | +0.1pp |
| SQuAD v2 | Headroom | 11.8% | −4.1pp |
| SQuAD v2 | Bear (TheTokenCompany) | 9.0% | −1.6pp |
| SQuAD v2 | LLM-Lingua (MSFT) | 62.5% | −19.4pp |
| GPQA Diamond | GPTZIP | 21.3% | +1.2pp |
| GPQA Diamond | Headroom | 12.2% | −2.5pp |
| GPQA Diamond | Bear (TheTokenCompany) | 5.8% | 0.0pp |
| GPQA Diamond | LLM-Lingua (MSFT) | 57.7% | −6.1pp |
| FinanceBench | GPTZIP | 16.5% | +0.9pp |
| FinanceBench | Headroom | 26.3% | −8.2pp |
| FinanceBench | Bear (TheTokenCompany) | 1.5% | −6.4pp |
| FinanceBench | LLM-Lingua (MSFT) | 60.3% | −21.9pp |
| HumanEval | GPTZIP | 14.0% | +2.0pp |
| HumanEval | Headroom | 14.7% | −6.0pp |
| HumanEval | Bear (TheTokenCompany) | 9.8% | +2.0pp |
| HumanEval | LLM-Lingua (MSFT) | 61.7% | −39.0pp |
| LongBench v2 | GPTZIP | 28.9% | +5.0pp |
| LongBench v2 | Bear (TheTokenCompany) | 9.8% | −3.2pp |
Method comparison
GPTZIP is the only lossless compression across use-cases.
| Benchmark | Uncompressed | GPTZIP | Headroom | Bear (TheTokenCompany) | LLM-Lingua (MSFT) |
|---|---|---|---|---|---|
| CoQA359/500 records | 76.6%Accuracy | 0.0ppLossless 76.6% accuracy13.9% median19.2% max | −21.6pp 55.0% accuracy51.8% median77.3% max | −0.1pp 76.5% accuracy13.1% median25.9% max | −18.5pp 58.1% accuracy63.0% median66.0% max |
| SQuAD v2366/500 records | 60.1%Accuracy | +0.1ppLossless 60.2% accuracy12.3% median22.4% max | −4.1pp 56.0% accuracy11.8% median51.1% max | −1.6pp 58.5% accuracy9.0% median23.1% max | −19.4pp 40.7% accuracy62.5% median67.6% max |
| GPQA Diamond198/198 records | 29.3%Accuracy | +1.2ppLossless 30.5% accuracy21.3% median70.7% max | −2.5pp 26.8% accuracy12.2% median95.4% max | 0.0pp 29.3% accuracy5.8% median19.8% max | −6.1pp 23.2% accuracy57.7% median82.7% max |
| FinanceBench113/150 records | 84.1%Accuracy | +0.9ppLossless 85.0% accuracy16.5% median40.0% max | −8.2pp 75.9% accuracy26.3% median83.8% max | −6.4pp 77.7% accuracy1.5% median14.1% max | −21.9pp 62.2% accuracy60.3% median78.1% max |
| HumanEval100/100 records | 58.0%Accuracy | +2.0ppLossless 60.0% accuracy14.0% median34.2% max | −6.0pp 52.0% accuracy14.7% median64.6% max | +2.0ppLossless 60.0% accuracy9.8% median33.6% max | −39.0pp 19.0% accuracy61.7% median70.8% max |
| LongBench v225/6 records | 42.1%Accuracy | +5.0ppLossless 47.1% accuracy28.9% median41.8% max | Not available | −3.2pp 38.9% accuracy9.8% median21.8% max | Not available |
Workload fit
Compression depends on the work you send.
Long contexts and short questions do not compress the same way. These rows use the closest measured public benchmark instead of one number for every workload.
| Workload | Public proxy | Median removed | Accuracy change |
|---|---|---|---|
| Long-context documents | LongBench v2 | 28.9% | +5.0pp |
| Financial and regulatory analysis | FinanceBench | 16.5% | +0.9pp |
| Code context | HumanEval | 14.0% | +2.0pp |
| Conversational and extractive QA | CoQA · SQuAD v2 | 12.3–13.9% | 0.0pp to +0.1pp |
| Expert reasoning | GPQA Diamond | 21.3% | +1.2pp |
Compression speed
330-token measured prompt
GPTZIP
30 ms
Headroom
200 ms
Bear (TheTokenCompany)
250 ms
LLM-Lingua (MSFT)
1200 ms
Private prompt compression
Compress private prompts safely.
The masking method runs on your browser, computer, or server. GPTZIP gets the masked text—not the original text. Drag the slider to see the live example.
Original input: Customer support case for Aurora Health: Maya Chen reported a duplicate charge on order 483901. The support agent must verify the billing event, preserve the refund policy, and answer whether the customer can keep the replacement device. Customer support case for Aurora Health: Maya Chen reported a duplicate charge on order 483901. The support agent must verify the billing event, preserve the refund policy, and answer whether the customer can keep the replacement device. Masked text: Client help type in Dawn Wellness: Alice David covered some twin care at club 145302. Any help factor need control each request issue, save each payment insurance, but reply house each buyer may hold each switch twist. Client help type in Dawn Wellness: Alice David covered some twin care at club 145302. Any help factor need control each request issue, save each payment insurance, but reply house each buyer may hold each switch twist.
Drag to compare original and masked text.
What leaves your device
The original data does not go to GPTZIP.
The customer-controlled device keeps the original text and the information needed to restore it. Only masked text goes to the GPTZIP service.
02
GPTZIP server
Receives only masked text and returns compressed masked output.
Masked text only
03
Compressed output
A shorter masked request comes back to your device.
68 tokens sent
04
Restore in the browser
The browser puts the original terms back into the user-facing text.
Restored locally
88
masked input
68
sent request
20
tokens saved
Your device restores the original terms after the GPTZIP step.
01 / BrowserOriginal inputReadable terms stay local.
Customer support case for Aurora Health: Maya Chen reported a duplicate charge on order 483901. The support agent must verify the billing event, preserve the refund policy, and answer whether the customer can keep the replacement device. Customer support case for Aurora Health: Maya Chen reported a duplicate charge on order 483901. The support agent must verify the billing event, preserve the refund policy, and answer whether the customer can keep the replacement device.Local only
For developers
Less tokens. More LLM quality.
Add GPTZIP before the LLM call or use GPTZIP Proxy. You do not need to change models or rebuild your harness.
- One API step for apps you own
- OpenAI-compatible proxy path for supported clients
- Before-and-after data you can inspect
For enterprise
Lower the token bill. Keep control of the rollout.
Start with one team, measure the savings, and expand after the results pass your review. Private Mode keeps the original text away from the GPTZIP service.
- Savings tied to measured token reduction
- Client-side masking for sensitive text
- API or proxy rollout by team and workflow
FAQ
What does GPTZIP do?+
GPTZIP shortens eligible prompts before they reach an LLM. A smaller request can use fewer input tokens, cost less, and take less data through the model request.
What does “lossless” mean here?+
GPTZIP checks whether the evaluated task result stays the same after compression. The benchmark section shows the before and after token count beside the checked result. You should still test GPTZIP on your own prompts before a broad rollout.
How fast is GPTZIP?+
In the current 330-token speed comparison, the GPTZIP compression pass takes 30 ms. The stated LLMLingua2 pass takes 1200 ms. Results change with prompt length, mode, hardware, and network setup.
How does GPTZIP lower AI costs?+
LLM providers usually charge for input tokens. When GPTZIP removes eligible input tokens, the billable input can be smaller. Your actual savings depend on the model price, the prompts you send, and the compression measured on your traffic.
How do developers add GPTZIP?+
Use the GPTZIP API before the model call, or point a supported OpenAI-compatible client to GPTZIP Proxy. The API fits apps that own the request code. The proxy fits tools that let you change the base URL.
Does GPTZIP work with LLM harnesses?+
GPTZIP works with most harnesses that support a normal API step or an OpenAI-compatible base URL. Exact support depends on how the harness lets you configure providers and proxies.
What is GPTZIP Private Mode?+
Private Mode runs the masking method in the customer’s browser, computer, or server. It changes selected original terms before the compression request. GPTZIP receives the masked version, while the information needed to restore the terms stays on the customer-controlled device.
Does Private Mode hide data from my LLM provider?+
Private Mode keeps the original text away from the GPTZIP compression service. It does not automatically change what your chosen LLM provider receives later in the workflow. That depends on how your team routes the final request.
Does GPTZIP replace prompt caching or RAG?+
No. Prompt caching lowers the cost of repeated prompt prefixes. RAG selects context. Model routing selects a model. GPTZIP makes eligible prompt text smaller. Teams can use these tools together.
How much setup does GPTZIP need?+
For the API, add one compression call before the LLM request. For GPTZIP Proxy, configure a supported client to use its compatible base URL. Start with a few real prompts, check the result, and then expand.
Next request
Cut the token bill, not the AI work.
Get access to the GPTZIP API, GPTZIP Proxy, and Private Mode.