An Economical Mannequin with Flagship Efficiency

GPT-6.1 Terra is designed for coding and sensible work, however how effectively does it deal with duties past writing code? From spreadsheets and displays to calendar information, these jobs require greater than a convincing response. The output must be correct, editable, and helpful after the dialog ends.

I examined GPT-6.1 Terra on 5 duties: analyzing a income workbook, repairing a Python script, making a lesson from a analysis paper, scheduling classes throughout time zones, and getting ready an audio file. On this article, I’ll stroll by means of the outcomes and look at how higher Terra is from its predecessors.

What’s GPT-6.1 Terra?

GPT-6.1 Terra is a cheap OpenAI mannequin designed for coding and sensible work. It accepts textual content and pictures, and may use instruments in supported environments to create and edit information, run code, and work by means of multi-step duties.

In ChatGPT Work and Codex, that may imply constructing a spreadsheet, repairing a Python script, or getting ready a presentation. Its file-handling capabilities rely upon the instruments obtainable within the surrounding atmosphere. For instance, processing an audio file with code doesn’t imply the mannequin can take heed to it natively.

Key options of GPT-6.1 Terra embody:

• Coding help: Writes, debugs, and explains code.
• Textual content and picture enter: Works with written directions and visible materials.
• Software use: Can run code and use obtainable instruments to finish duties.
• File dealing with: Can create and edit paperwork, spreadsheets, and displays in supported environments.
• Multi-step work: Can perform duties that contain a number of phases, similar to inspecting information, making adjustments, and checking outcomes.
• Economical operation: Designed for sensible work the place price effectivity issues.

GPT-6.1 Terra Arms-On Assessments

I examined GPT-6.1 Terra throughout 5 sensible duties involving spreadsheets, Python, displays, calendars, and audio. For every job, I checked the output towards the immediate and checked out whether or not the ensuing file or code was prepared to make use of.

1. Construct an Excel Income Audit With Working Formulation

The ten-row order ledger consists of refunded and pending orders, a precise duplicate of AV105, and a paid order, AV107, and not using a value. The deliverable was a workbook somebody may open and edit, with the audit preserved alongside the figures.

Immediate

Flip orders.csv into an editable Excel workbook, not a pasted desk. Preserve an untouched Uncooked sheet, a Cleaned sheet that deduplicates precise orders and consists of solely paid orders, and a Abstract sheet. Calculate net_revenue with Excel formulation after share reductions, flag lacking costs with out guessing, and make the entire a method. Add an audit of excluded and duplicate IDs. Don’t invent a foreign money. Save orders_audit.xlsx and provides me the obtain plus a word beneath 90 phrases.

What Occurred

The workbook has Uncooked, Cleaned, and Abstract sheets. Uncooked preserves all ten supply rows; Cleaned retains six paid orders, together with unresolved AV107. Income cells include IF formulation, and the Abstract makes use of formulation for the identified complete, retained depend, and missing-price warning.

The method outcomes independently checked out: identified income is 4,545.20, with AV105 counted as soon as. AV102 and AV109 are excluded as refunded, and AV104 as pending. AV107 stays clean and flagged. The whole is incomplete, appropriately.

One restrict: the cleaned supply columns are a copied snapshot, not a reside filter of Uncooked. Modifying a cleaned row recalculates its income; including new supply rows would require refreshing the cleansing step. Helpful workbook, with a transparent upkeep boundary.

2. Restore a Python Perform and Run Its Edge Instances

Subsequent got here overdue_ids, a perform that returns sorted IDs for open tickets whose due time has handed. The unique code used <=, which incorrectly included tickets due precisely now, and accessed due dates with out dealing with lacking or malformed values.

Immediate

Restore ticket_triage.py with a minimal patch so overdue_ids follows its docstring. Truly run checks for Z timestamps, differing UTC offsets, equal-to-now, lacking and malformed due dates, closed tickets, sorted IDs, and no enter mutation. Present the whole corrected perform in your response, then report the precise test lead to beneath 50 phrases. State how timezone-naive dates are handled. Don’t declare a check handed except you ran it.

Code response, half 1. The perform begins right here and explicitly paperwork its UTC assumption for timestamps and not using a timezone.

What Occurred

Terra modified the comparability to due < now, normalized Z timestamps, and caught lacking or invalid due values. The app’s exercise confirmed a command run for the checks, and its closing response reported the listed instances passing.

I downloaded the corrected Python file and ran impartial checks for various offsets, equal instants, invalid and lacking due dates, closed tickets, sorting, empty enter, and preservation of the unique knowledge. The patched perform handed.

The primary implementation selection is price checking: Terra treats timezone-naive due dates and present instances as UTC, and provides that coverage to the docstring. That’s specific and testable, however an actual help crew would possibly require native time as an alternative. The perform nonetheless assumes the anticipated standing and ID fields. This was a scoped restore, not a assessment of a complete ticketing system.

3. Make a Brief Self-Consideration Lesson From a Analysis PDF

An evidence that matches in a chat is just not routinely a usable lesson. I equipped the 15-page Consideration Is All You Want paper and requested a three-slide PowerPoint for a Python learner. The file wanted editable content material and speaker notes, with the arithmetic preserved.

Immediate

Use the connected Consideration Is All You Want paper to make an editable three-slide PowerPoint lesson for somebody who is aware of Python however is new to Transformers. Slide 1: clarify queries, keys and values. Slide 2: give the precise scaled dot-product consideration equation, outline d_k, and clarify the scaling. Slide 3: distinguish encoder self-attention from masked decoder self-attention and clarify why positional data is added. Preserve every slide beneath 55 seen phrases, with detailed speaker notes that time to the related paper sections. Use textual content, equations and easy tables solely; no illustrations or generated photos. Save attention_lesson.pptx. Test there are precisely three slides and report any issues in beneath 80 phrases.

Terra provides the PowerPoint and explicitly says it couldn’t test visible rendering.

What the Deck Incorporates

The downloaded deck comprises precisely three slides, with 47, 43 and 46 seen phrases respectively, together with the operating header and web page counter. Its textual content stays editable. Speaker notes clarify the fabric in additional element and determine the related sections of the paper.

I inspected all three slides within the app viewer. Their textual content was readable and the equation remained intact. Terra’s personal response didn’t declare that visible test: it reported that rendering was unavailable. The result’s a compact educating help; it doesn’t cowl multi-head consideration or a full Transformer implementation.

4. Produce Calendar Occasions Throughout 4 Speaker Timezones

The workshop file offers 4 talks, one room, prerequisite classes, ten-minute breaks, and availability expressed in London, Kolkata, Singapore and New York time. The occasion date is 3 November 2026. I wished the earliest possible timetable and an ICS export with unambiguous UTC timestamps.

Immediate

Plan the workshop from event_schedule.json. Discover the earliest potential end whereas respecting every speaker’s precise timezone, availability, conditions, one-room restrict and breaks. Create workshop.ics with one occasion per session and UTC timestamps ending in Z, following each calendar requirement within the file. Don’t ship invites or write to a calendar. Return the file plus a compact schedule in London and Kolkata time. Clarify in beneath 80 phrases why the schedule finishes as early as potential.

The whole timetable, calendar obtain and lower-bound rationalization, with Terra’s mannequin label seen.

Schedule and File Checks

Terra scheduled A–C–B–D from 09:00 to 12:15 London time, equal to 14:30–17:45 in Kolkata. The 4 classes use 165 minutes; three ten-minute breaks carry the minimal elapsed time to 195 minutes. Unbiased enumeration of all 24 potential orders confirmed the consequence.

The ICS comprises 4 occasions with distinctive identifiers, right begin and finish values in UTC, and descriptions containing the audio system and each show timezones. Its line endings and folded traces additionally handed format checks. I inspected the export; it was not imported into an exterior calendar.

5. Put together a Stereo WAV for a Mono Audio Pipeline

For a numerical file-processing check, I used an eight-second stereo WAV with identified 440 Hz and 880 Hz indicators, a DC offset, and one second of silence at every finish. The identified sign lets me test the transformation exactly. That is audio processing by means of instruments, not a listening or transcription check.

Immediate

Put together the uploaded stereo WAV for a mono audio pipeline. Preserve precisely enter seconds 1–7. Common the 2 channels, take away the DC offset, resample to 16,000 Hz, and normalize the pattern peak to -1 dBFS earlier than making use of 50 ms linear fade-in and fade-out. Export mono PCM 16-bit WAV as prepared_audio.wav. Do the processing, then confirm period, pattern price, channels, bit depth, peak and remaining DC. Return the obtain and a report beneath 100 phrases. This can be a signal-processing job; don’t transcribe or describe what it appears like.

The precise WAV export and Terra’s reported measurements.

Output Measurements

The returned WAV comprises precisely 96,000 mono samples: six seconds at 16,000 Hz, saved as 16-bit PCM. Its measured peak is roughly −1.00 dBFS, and its imply offset is 0.000000373 of full scale. Each endpoints are zero after the fade.

I independently reconstructed the requested transformation from the enter. The output carefully matched that reference, with normalized root-mean-square error of about 0.022%. That checks the precise samples, past the file label. It doesn’t set up speech-recognition high quality or perceptual audio enhancement.

GPT-6.1 Terra benchmarks

OpenAI examined GPT-6.1 Terra on skilled pc duties, visible reasoning, factuality and coding. Listed here are the outcomes from its launch, adopted by what every check measures. These are OpenAI’s printed evaluations, separate from our hands-on assessments.

Brokers’ Final Examination

This benchmark covers skilled pc duties, together with deciphering drawings and reviewing animation. At medium reasoning, Terra scored 57.6%, matching GPT-6 Astra and exceeding GPT-5.6 Terra’s 42.6%. The chart estimates $0.29 per job for Terra and $4.82 for Astra at that setting.

MMMU-Professional

MMMU-Professional evaluates visible reasoning. Terra outperformed GPT-5.6 Terra throughout reasoning settings. At most effort, it scored 85.5%, in contrast with 86.3% for GPT-6.1 Sol and 86.7% for GPT-6 Astra.

OSWorld 2.1

OSWorld assessments multi-step work in desktop functions. OpenAI stories that Terra at medium effort surpassed GPT-5.6 Terra’s greatest rating at a fraction of the fee per job. The chart reveals the offline set.

Factual errors

On OpenAI’s inner set of error-inducing prompts, Terra’s low-effort error price was 16.7%, down from 53.6% for GPT-5.6 Terra. At most effort it was 4.3%. Decrease is healthier; this chosen set doesn’t signify an error price throughout bizarre requests.

A Fastened Token Funds

For Terra’s Commonplace API text-token charges, a million recent enter tokens price $1, and the identical quantity of output prices $5. To offer the economic system a concrete unit, I priced a workload containing 80,000 uncached enter tokens and 20,000 complete output tokens.

The commentary was made on Codex utilizing the OpenAI API.

Conclusion

GPT-6.1 Terra dealt with a spread of sensible duties in these assessments, from fixing Python code and constructing a spreadsheet to making a lesson, calendar, and audio file. Its worth lies in producing information you possibly can examine, edit, and use past the dialog.

The outcomes additionally present why assessment issues. The workbook left AV107’s value unresolved, whereas the Python patch documented UTC as its default. These particulars are seen within the information, so customers can test them earlier than counting on the output.

For on a regular basis coding and file-based work, Terra seems to be like a helpful choice when paired with the precise instruments. These assessments are a small pattern, although, and don’t set up the way it will carry out on bigger duties or whether or not it is going to be cheaper for a selected job.

For context on the broader lineup, see our GPT-6 Astra explainer. The OpenAI Dot walkthrough explores the app expertise round tool-driven work.

Regularly Requested Questions

Finding out, evaluating, and explaining AI techniques for over 6 years.

“𝘖𝘯𝘤𝘦 𝘮𝘦𝘯 𝘵𝘶𝘳𝘯𝘦𝘥 𝘵𝘩𝘦𝘪𝘳 𝘵𝘩𝘪𝘯𝘬𝘪𝘯𝘨 𝘰𝘷𝘦𝘳 𝘵𝘰 𝘮𝘢𝘤𝘩𝘪𝘯𝘦𝘴 𝘪𝘯 𝘵𝘩𝘦 𝘩𝘰𝘱𝘦 𝘵𝘩𝘢𝘵 𝘵𝘩𝘪𝘴 𝘸𝘰𝘶𝘭𝘥 𝘴𝘦𝘵 𝘵𝘩𝘦𝘮 𝘧𝘳𝘦𝘦. 𝘉𝘶𝘵 𝘵𝘩𝘢𝘵 𝘰𝘯𝘭𝘺 𝘱𝘦𝘳𝘮𝘪𝘵𝘵𝘦𝘥 𝘰𝘵𝘩𝘦𝘳 𝘮𝘦𝘯 𝘸𝘪𝘵𝘩 𝘮𝘢𝘤𝘩𝘪𝘯𝘦𝘴 𝘵𝘰 𝘦𝘯𝘴𝘭𝘢𝘷𝘦 𝘵𝘩𝘦𝘮.” — 𝖥𝗋𝖺𝗇𝗄 𝖧𝖾𝗋𝖻𝖾𝗋𝗍, 𝖣𝗎𝗇𝖾