A/B test results

Claude Sonnet 5 vs GPT-5.6 Terra
for A/B test results

This page compares the two models on one job: reading a pasted A/B test results table and judging whether the lift is real or just noise. It ends with a fair way to test both on your own experiments.

Sep 8, 2026 · 9 min read

The bottom line
Terra edges it but only provisionally

GPT-5.6 Terra is the safer first choice for turning a pasted A/B test table into a verdict, based on stronger performance reading and calculating from quantitative material. That evidence is indirect, since no benchmark tests either exact model on A/B-test significance itself. Claude Sonnet 5 stays cheaper and holds up better against harmless table changes, but independent testing found it keeps answering even when a table's evidence is missing or contradictory.

That split shows up in three places: a chart-reading benchmark where Terra scored far higher1, a finance benchmark where the two are close enough to call a tie2, and a structured-data reliability test that found Sonnet 5 rarely stopped when a table couldn't support an answer3. None of them tests A/B-test significance directly, which is the real gap in the public evidence.

For a live workflow, the safer split is stage by stage: use Terra for the first parsing pass on a messy or dense table, let either model write the plain-language explanation once the numbers are set, and always run the actual significance calculation through a deterministic statistics function rather than trusting either model's arithmetic. Save analyst review for any launch, medical, financial or regulatory decision.

Who this is for
Which analysts this fits

Start with Terra01

Growth and product analysts

You paste a variant's results table and need a first read fast. Terra tested stronger at extracting and calculating from dense quantitative material, so it is the safer opening pass before a deterministic check.

Add a stats tool02

Experimentation platform teams

You run this at volume. Give either model a validated statistics function to call instead of trusting its own arithmetic, and keep the model's job to reading and explaining, not calculating.

Lean on Sonnet 503

Teams on tight formats

Your reporting template changes often but the underlying fields stay complete. Sonnet 5 held its answers steadiest when valid data was reordered or reformatted in independent testing.

Add human review04

High-stakes launch calls

A ship or no-ship call rides on the result. Neither model has a published record proving it knows when to abstain, so an analyst should check the extracted counts and the test before anyone acts.

What we compared
Reading the table not the app

This page compares the two models through their API in one neutral setup, not one model wrapped inside one app against the other wrapped inside a different one.

The parts that matter for this job are reading the counts and rates correctly, picking a defensible test, calculating the lift and the interval right, telling a real effect from a chance blip, and knowing when the table is simply missing a field the analysis needs. Official docs come first, then independent benchmarks that test extraction, arithmetic and evidence handling under matched conditions.

Tool rows are left out of the spec table on purpose. A spreadsheet add-on, a code sandbox or a file-upload feature belongs to the app built around the model, not to the model itself, and the same model can behave differently depending on which one it sits inside. The one exception is tool or function calling, since handing extracted counts to a validated statistics function is a model-level capability both APIs support95.

Specs at a glance
The numbers this job needs

Both context windows are far bigger than one pasted table needs. Accuracy and evidence handling matter more than context size here.

Spec
Claude Sonnet 5
GPT-5.6 Terra
Why it matters
Context window
1,000,000 tokens
1,050,000 tokens
Both are far more than a pasted table needs. Extra room only helps if a long experiment brief comes attached45
List price
$2 in / $10 out per million tokens
$2 in / $12 out per million tokens
Input is priced the same. Sonnet 5 is cheaper on the generated explanation, which is mostly output tokens45
Long-input pricing
Standard rate across the full window
$4 in / $18 out per million once a request passes 272,000 input tokens
Rarely triggered by one short table, but it changes the price the moment a long brief rides along5
Max output
128,000 tokens
128,000 tokens
Matched, and more than either model needs for a table read-out and its explanation45
Reasoning control
Adaptive thinking with a settable effort level
Reasoning effort from none through max
Turn effort up for a hard audit and down for routine volume work on either model45
Tool calling
Structured JSON output and tool use
Structured outputs and function calling
Both can hand extracted counts to a shared statistics function instead of doing the maths in the model95

Figures from Anthropic and OpenAI documentation, checked September 8, 2026. The two vendors price and tokenize differently, so treat any cross-model cost comparison as directional, not exact.

Head to head
Where each model's edge sits

This is the main analysis: which model has the edge on each part of reading a results table and deciding if a lift is real, and what backs it up.

Job
Better choice
Why the edge exists
Best evidence
Reading a dense quantitative table
GPT-5.6 Terra
A chart-reading benchmark tests extraction and reasoning over professional quantitative displays under matched conditions. It uses charts rather than plain pasted text, so the result is directional for table reading, not an exact match.
Terra scored 34.0 versus Sonnet 5's 16.6 at maximum effort across 100 tasks1
Routine quantitative calculation
Statistical tie
An independent finance benchmark scores both close together, and its own method calls overlapping confidence intervals a tie rather than a win.
87.7% of tasks passed for Terra against 84.9% for Sonnet 5, with overlapping intervals2
Recognizing missing or contradictory evidence
Terra, provisionally
This is a risk call rather than a measured Terra win, since Terra was not tested on this exact question. A structured-data benchmark found Sonnet 5 usually kept calculating on tables built to be unanswerable.
Sonnet 5's average reliability score was 13.58%, with 80 to 84% no-refusal rates on unanswerable table questions3
Surviving harmless table reformatting
Claude Sonnet 5
The same benchmark reordered valid data and added irrelevant rows around it. Sonnet 5 held its answers steady better than any model tested, though Terra was not included in this part of the test.
Sonnet 5 recorded the lowest average attack-success rate, at 9.10%3
Getting the significance number right with a tool
Tie
Both APIs publish tool and function calling, so a validated statistics function can take over the arithmetic once the model extracts the counts. The remaining difference is table reading, already covered above.
Both vendors document tool calling for this exact use95
Cost for one short pasted table
Claude Sonnet 5
Input pricing is identical. A table explanation is mostly output tokens, and Sonnet 5's output rate is lower.
Sonnet 5 lists $10 per million output tokens against Terra's $1245
Calibrated language versus a confident answer
No proven winner
Sonnet 5 has the direct negative evidence above on unsupported tables. Terra has no equivalent public, exact-version calibration result. Overconfidence shows up broadly across modern models, which points to a real risk without deciding between these two.
Nominal 99% intervals covered the true answer only 65% of the time on average across tested models7

Better-choice calls map to what the sources actually tested. No public benchmark grades either exact model on A/B-test significance itself, so the first three rows are the closest available evidence, not a direct test of this task.

How to test
A fair test on your own results table

Build a small set from three to five real experiments, then add adversarial variants: a clear lift, a small-sample lift, a borderline result, a table missing a field, and a multi-metric test that needs a multiple-comparison warning.

Sample01

Pick three to five experiments

Cover a large clearly significant lift, the same apparent lift with a small sample and a wide interval, a borderline result near your threshold, a table missing denominators or unit definitions, and a multi-variant test.

Prompt02

Use one shared prompt

Give both models identical source text and ask for the metric type, the absolute lift, the relative lift, the confidence interval, the p-value and a plain-language caveat about what the result does and does not prove.

Setup03

Match effort and add the tool

Set the same reasoning effort for both, and hand each model the same deterministic statistics function to call rather than letting it do the arithmetic alone. Calculate the reference answers yourself first, in R, Python or your experimentation platform.

Scoring04

Score before any edits

Do not edit an output before scoring it. Check the extracted counts, the chosen test, whether it separated failure to reject from proof of no effect, and whether it flagged missing information. For a real launch call, use blind review by at least two analysts.

What the evidence shows
No direct benchmark so read the parts

No published benchmark grades these two exact models on A/B-test significance. Here is what each nearby source actually measures, and how much weight to put on it.

Source
What it measures
What it suggests
How to weigh it
Chartography
Reading and reasoning over professional charts, 100 tasks with repeated trials
Terra carries much lower risk when extracting figures from a dense quantitative display
Directional for table reading. Its input is charts, not plain pasted text, so it is not an A/B-test score1
Finance Benchmark
Verifiable numerical, risk and code-execution tasks with seeded answers and confidence intervals
Terra's point estimate is a little higher, but the intervals overlap enough that the benchmark itself calls it a tie
Close to the arithmetic component of this task, but it has no dedicated A/B-testing suite2
TrustDABench
Whether a model stops on missing or conflicting table evidence, and how it handles harmless reformatting
Sonnet 5 followed a valid data path well but rarely stopped when no valid path existed. Terra was not tested
The benchmark most relevant to the confident-regardless question, though only one of these two models is in it3
StatABench and FermiEval
General statistical method choice and confidence-interval calibration across many models
Models broadly choose the wrong method too often, and stated confidence intervals run overconfident
Neither compares these two exact models, so treat both as background warnings rather than a winner call67

Single-tester and community write-ups are a secondary signal only. None of the sources reviewed for this page treats A/B-test significance as its own benchmark task.

How to prompt each one
Same table needs a different contract

Terra does best with an explicit analysis contract and a higher effort setting for a one-off audit. Sonnet 5 needs the abstention rule stated as loudly as the calculation itself.

GPT-5.6 Terra works best with an explicit analysis contract at a higher reasoning effort for a one-off audit, or medium effort plus the shared statistics tool for volume work. Ask it to identify the metric type, numerator, denominator, unit of analysis and assignment design before it calculates anything, stop if a required field is missing, then call the supplied function and report the lift, interval, p-value and a plain caveat5.

Claude Sonnet 5 needs the abstention rule stated as clearly as the calculation itself. Tell it directly not to infer a missing sample size, event count, independence assumption or metric definition, and to output only a fixed marker plus the missing fields when the table cannot support one unique analysis. That instruction targets the exact unsupported-path behavior an independent benchmark documented for Sonnet 53.

A Claude Sonnet 5 prompt: a loud abstention rule

Before calculating anything, run an evidence-sufficiency check.
You must not infer sample sizes, event counts, independence,
test duration or metric definition.

If the table cannot support one unique analysis, output only
INSUFFICIENT_DATA and the missing fields.

If it can, call the supplied statistics function and explain
the result in cautious, non-technical language.

A GPT-5.6 Terra prompt: an explicit analysis contract

Read the table without assuming missing values. Identify metric
type, numerator, denominator, unit of analysis and assignment
design.

If any required field is absent, stop and ask for it.

Otherwise call compare_proportions, then report absolute lift,
relative lift, 95% CI, two-sided p-value, practical
interpretation and caveats. Do not describe p>0.05 as proof of
no effect.

Weak spots
Where the cleanup work still sits

Neither model is safe to trust unattended. The useful question is where each one adds cleanup work, and what to change in the prompt or the workflow.

Model
Weak spot
What it looks like
How to fix it
Claude Sonnet 5
May keep going on an unanswerable table
It can make a plausible assumption about a missing sample size or event count and still produce a confident number, and it reproduces that same assumption stably across reformatted versions of the table, which makes the wrong answer look reliable.
Add an evidence checklist and a fixed abstention output, and separate extraction, sufficiency checking, calculation and explanation into distinct steps. Anthropic reports this behavior improved relative to Sonnet 4.6, but that does not cancel this independent result310.
GPT-5.6 Terra
No published calibration result for this exact version
Its stronger extraction and calculation scores do not by themselves show whether it flags optional stopping, peeking, multiple metrics, a sample-ratio mismatch or a paired design.
Spell out each of those checks in the prompt and the tool schema, and reject any answer that skips listing its assumptions and the test it chose8.
Both
Can turn a p-value into a plain ship or no-ship answer
Either model can report only the p-value and skip the absolute lift, or confuse a relative lift with an absolute percentage-point change.
Require both effect measures, the interval, the decision threshold and a separate practical-significance line, and keep the actual number coming from deterministic code plus analyst review6.

Which one to choose
Start from whether a tool checks it

One question first. Will a deterministic statistics function calculate the result, or is the model reasoning and calculating alone? Then follow the branch that matches your situation.

Does a function calculate the result? No: missing-data check is critical No: high-stakes or compliance call Yes: lowest price or many short replies Yes: table is messy or dense Yes: format changes fields stay complete Terra plus abstention checklist Neither without human review Claude Sonnet 5 GPT-5.6 Terra Claude Sonnet 5 Verify every number yourself

A starting point, not a rule. Test on your own experiments before you commit.

Recommendations
Start from who checks the answer

If a validated statistics function will do the actual calculating, the model choice comes down to parsing and price. Pick Claude Sonnet 5 when you need the lowest output-token cost or you are running many short explanations, and when the table's format changes often but the underlying fields stay complete. Pick GPT-5.6 Terra when the incoming tables are messy, dense or hard to parse145.

If the model has to reason and calculate without a shared tool, GPT-5.6 Terra is the safer general default on the adjacent evidence, and it should run at high or max effort. Missing-data detection stays the open question either way, so give Terra a hard abstention checklist rather than trusting it to flag a gap on its own, since no public result shows it will37.

For a launch, medical, financial or regulatory decision, neither model replaces analyst review. Use whichever model reads faster as a first pass, then check the extracted counts, the chosen test and the interval against an independent calculation before anyone acts on it67.

One case sits outside all of this. A solo analyst running occasional checks with no team to share results with gets little from Playgram's shared workspace, built around team memory and multi-person usage controls, so a single model subscription is the simpler and cheaper choice for that one person. The setup below earns its keep once more than one person needs the same comparison and the same context.

Bottom line
A provisional edge for GPT-5.6 Terra

GPT-5.6 Terra is the better provisional choice for reading a pasted A/B test table and judging whether a lift is real. Its strongest advantage is quantitative extraction. Claude Sonnet 5 costs less and holds up better against harmless table changes, but it has the clearer documented tendency to keep answering when the table's evidence is not enough.

The important limit is that no direct public benchmark compares these exact models on A/B-test significance, and Terra has no published, exact-version result proving it knows when to stop. The benchmarks behind this page also differ in prompts, reasoning effort, available tools, graders and input type, and prices and behavior can change quickly, so treat every figure here as a starting point rather than a permanent fact123.

The safest final step is to test the shape of your own experiment data, not a generic prompt from the internet. A fair test needs the same setup for both models: the same table, the same prompt and the same place to run them, so the result reflects the models and not the tool around them. In practice that is harder than it sounds, since most teams end up running one model in one app and the other in a different one, on two separate subscriptions, which tilts the comparison before the first answer comes back. The cleaner the setup, the more the difference you see is really Claude Sonnet 5 vs GPT-5.6 Terra, and not just which one happened to be easier to reach that day.

Run this test in one workspace
Right here inside Playgram

That's the practical case for one steady setup, and it's also what makes the daily work faster. When both models sit in one workspace, you can paste the same table to each, compare the two write-ups side by side, and hand a partial analysis from one model to the other without pasting the table in again.

Playgram lets you run that exact test directly: paste an A/B test results table once, put it in front of the latest GPT and Claude models, and keep asking follow-up questions with either one without re-pasting the table or starting the analysis over for a second opinion.

The same memory carries across the team too, not just this one table, over the latest GPT, Claude, Gemini and Grok models and many more, all in one place11. The line-up is curated, so retired models are turned off and new ones are added as they ship.

Team memory

Shared across everyone and every model.

Project memory

Scoped to a campaign or document set.

Personal memory

Your own working style, kept private.

Fair pricing
Pay per usage, not per seat

Upgrade as needed, and only pay for what you actually use

Save ~17% with the annual plan

Pro

$50/ month

Perfect for small and medium teams

Unlimited users & infinite memory

Multi-LLM chats

Granular access control to models

EU data residency

Get started

Ultra

$200/ month

Best for large, growing teams

Unlimited users & infinite memory

Multi-LLM chats

Unlimited use of DeepSeek V4 Flash

Granular access control to models

Choose US or EU data residency

Get started

Enterprise

Get in touch

Unlimited Credits

For organizations with advanced needs

Unlimited users & SSO

Priority Support

Unlimited use of DeepSeek V4 Flash

Granular access control to models

Choose US or EU data residency

Book a call

30-days money back guarantee

Frequently asked
questions

No. Treat the model's number as a draft, not a verdict. Give it a validated statistics function to call for the actual calculation, since both APIs support tool calling, and recalculate the answer independently in R, Python or your experimentation platform before anyone ships on it.

Neither has a proven safe record here, but the risk is asymmetric. Independent testing found Claude Sonnet 5 usually kept answering a table built to be unanswerable, with a reliability score of 13.58% and no-refusal rates around 80 to 84%. GPT-5.6 Terra was not tested on the same benchmark, so its behavior on missing data is unproven rather than confirmed safe.

It can, and the two models differ here. Claude Sonnet 5 held its answers steadiest when valid data was reordered or surrounded by irrelevant rows in independent testing, recording the lowest attack-success rate of the models in that benchmark. GPT-5.6 Terra was not part of that specific test.

Not proven overall. It has a clear lead on a chart-reading benchmark and a small, statistically tied lead on a finance-calculation benchmark, but no public benchmark tests either exact model on two-proportion significance, confidence intervals or multiple-comparison correction directly. Terra's edge on this exact task is a provisional read from adjacent evidence, not a measured result.

Force an evidence check first. Ask it to name the metric type, the numerator, the denominator, the unit of analysis and the assignment design, and to stop and say what is missing rather than guessing. Only after that should it call a shared statistics function and report the lift, the interval, the p-value and a plain caveat about what the result does and does not prove.

No. Run the same prompt on both and compare the answers, or switch between them mid-conversation. You choose after reading both answers instead of guessing up front.

Related comparisons

Claude Sonnet 5 vs GPT-5.6 TerraClaude Sonnet 5 vs GPT-5.6 Terra for editing draftsClaude Sonnet 5 vs GPT-5.6 Terra for grant writingClaude Sonnet 5 vs GPT-5.6 Sol for policy translation

One table for both models
One place to check the maths

Paste the same A/B test results table into the latest GPT and Claude models, run the numbers through one shared statistics function, and keep the whole comparison in one place. Set it up in a minute.

Get startedSee the pricing