As of November 16, 2026, OpenAI’s general-purpose choices cannot be summarized as one “latest GPT.” Chat, ChatGPT Work, Codex and the API differ in models and permissions. Recent Wallet, voice and image updates also do not mean a new general reasoning model. We establish the current entry points, then use GPT-6.1 Terra to examine why vendor charts, independent experiments and user experience can disagree.
OpenAI model release notes · ChatGPT Work and Codex models guide · ChatGPT product release notes · ChatGPT Chat models and thinking-level help
We opened official documentation, original Artificial Analysis results, Bito’s own tests and public Reddit discussions. They establish product boundaries, measured results under particular conditions and personal experiences respectively. They are not substitutes for one another. The synthetic invoices, prompts and acceptance gate are our own method examples. Only the offline gate’s self-tests were executed: no model calls, account access, purchases or subscription-allowance measurements.
Identify the current ChatGPT surface and model family
In Chat, GPT-6 Sol powers Instant through Extra High; the Pro reasoning option uses GPT-6 Astra. Available settings depend on plan, workspace and rollout. Plus access to Astra in Work/Codex does not establish access to Pro reasoning in Chat. Recommendations for GPT-6.1 Sol and Terra come from Work/Codex documentation and should not be copied into a Chat model-picker list.
OpenAI model release notes · ChatGPT Work and Codex models guide · ChatGPT product release notes · ChatGPT Chat models and thinking-level help
Official pages are not fully synchronized. The October 7 announcement says Free/Go will use GPT-6 Luna, while the detailed Chat help we examined still describes GPT-5.6 Luna as the default and Think model and retains explicit older selections. We do not invent a universal switchover date. Record the actual picker, plan, client and date. The table calls this a Luna-family migration rather than promising an account has migrated.
Terra launched October 26 for Work, Codex and the API, not Chat. Its lack of UAE inference residency concerns where computation runs, not a blanket restriction based on a user’s physical location. API use and subscriptions have separate billing; a subscription also does not guarantee every tool, region, administrator permission or preview entitlement.
| Surface | Models in current documentation | Boundary |
|---|---|---|
| Chat: GPT-6 | Sol; Astra for eligible Pro reasoning | Plan/admin/rollout affect settings; no6.1 Terra in Chat |
| Chat: Free/Go and older selections | Luna-family migration; detailed help still says5.6 | Announcement and help differ; record the actual picker |
| Work / Codex | Astra,6.1 Sol,6.1 Terra,Luna | Tools, client and workspace must support them; not Chat access |
| API | Specific model ID and API billing | Separate ledger; no fixed subscription-to-request conversion |
OpenAI model release notes · ChatGPT Work and Codex models guide · ChatGPT product release notes · ChatGPT Chat models and thinking-level help
Separate vendor evidence, independent tests and community reports
Official announcements and system cards establish intended use, configuration and vendor-run test conditions. Independent evaluators remeasure models in their own datasets and harnesses. Community posts establish that someone reported an experience. Having all three is useful, but repeated coverage of the same chart is not several independent experiments.
OpenAI GPT-6.1 Terra announcement · GPT-6.1 Terra system-card addendum · Artificial Analysis original Terra results · Bito independent coding tests · Reddit: unfinished-work early-stop report
Write the comparison as model version × effort × product surface × tools/harness × task × scoring. Change any component and the numbers may no longer be directly comparable. Faster output can mean faster service, less work or earlier stopping. A better score can use a larger reasoning budget. A confident completion message is still not evidence of a saved result.
| Material | Supports | Does not directly establish |
|---|---|---|
| Official release/system card | Dates, intended use and vendor-test conditions | Identical success for every user |
| Original independent evaluation | Results for its dataset, harness and budget | Unconditional cross-product or cross-task superiority |
| Direct community account | Reported behavior and troubleshooting hypotheses | Representative samples, verified routing or a failure rate |
OpenAI GPT-6.1 Terra announcement · GPT-6.1 Terra system-card addendum · Artificial Analysis original Terra results · Bito independent coding tests · Reddit: unfinished-work early-stop report
What the vendor’s strengths establish—and their limits
OpenAI positions Terra for visual understanding, computer use and clear writing. Its announcement reports 50.3% on Terminal-Bench 4.0 at max versus 21.5% for the previous Terra. On an internal error-inducing prompt set at low effort, the error rate falls from 53.6% to 16.7%. These are vendor results on specific evaluations, not a promise of the same reduction on every real question.
OpenAI GPT-6.1 Terra announcement · GPT-6.1 Terra system-card addendum
The system card explicitly notes that research or API results may differ from production ChatGPT because of system prompts, tools and effort. We therefore do not expand “cheaper” or “faster” into a guarantee for every ChatGPT task. Choose work with known answers or inspectable artifacts first, then test whether the claimed strength becomes successful delivery.
| Evaluation | Reported values | Conditions and limits |
|---|---|---|
| Terminal-Bench4.0 | Terra50.3%; previous Terra21.5% | max; vendor-run coding evaluation |
| Internal error-inducing prompts | low error rate53.6%→16.7% | Internal set, not the internet-wide error rate |
| Vision/computer use | Vendor reports relative improvements | Match task, effort, tools and baseline |
OpenAI GPT-6.1 Terra announcement · GPT-6.1 Terra system-card addendum
AA changed index versions: do not splice old scores into new ones
The current Artificial Analysis Terra page describes Max and shows 51 on Intelligence Index v5.0.1, an average $0.96 per task and 13.9 minutes per task. The index incorporates nine evaluations. A score of 51 is not 51% accuracy, and 13.9 minutes is not ordinary chat latency. These figures describe that test mixture and reasoning setting.
Artificial Analysis original Terra results · Artificial Analysis index methodology
Early coverage often quotes 48 and $0.40 from an older version. A new methodology, dataset composition or weighting changes the comparison. Do not subtract the old number from today’s and claim the model gained three points or suddenly became more expensive. Preserve access date, version, settings and the original page. Mark results noncomparable when matching conditions cannot be established.
| Metric | Current page | Interpretation |
|---|---|---|
| Intelligence Index | v5.0.1;9 evaluations;51 | Composite index, not51% accuracy |
| Average task cost | $0.96 | Estimated cost for index tasks, not a subscription price |
| Average task time | 13.9minutes | Includes evaluation reasoning, not everyday chat latency |
Artificial Analysis original Terra results · Artificial Analysis index methodology
Bito’s coding test: speed gains and three-point noise
Bito’s own coding comparison, updated November 11, uses default medium in each vendor’s coding agent. Two difficult five-turn tasks receive 75% weight and one-shot questions 25%; Terra had not entered its twenty-turn test. It reports Terra at 69.0, $0.40 and 4.4 minutes per difficult session. These are not AA’s tasks, costs or a general success rate.
Haiku scores 67.4, a 1.6-point gap; the author reports approximately three points of session noise, so this is a close result within that experiment. The gap to Opus is larger, but harness, sample and judge still limit transfer. Bito sells coding tools: independence from OpenAI does not mean absence of commercial interests or establish a universal coding verdict.
| Model | Weighted score | Cost/hard session | Minutes/hard session |
|---|---|---|---|
| GPT-6.1 Terra | 69.0 | $0.40 | 4.4 |
| GPT-6.1 Sol | 74.8 | $2.06 | 22.7 |
| Claude Haiku5.5 | 67.4 | $0.69 | 8.1 |
| Claude Sonnet5.5 | 65.7 | $1.28 | 8.6 |
| Claude Opus5.5 | 85.0 | $5.40 | 19.0 |
Why community reactions diverge
A direct discussion reports Terra ending two authorized workflows unfinished on macOS through T3 Code. The author also discloses conflicting older instructions, truncated tool output and unverified per-turn effort. It supplies a test hypothesis about premature stopping and local blockers becoming global ones. It does not measure a general Terra failure rate or eliminate harness effects.
Reddit: unfinished-work early-stop report · Reddit: original Terra release discussion · Reddit: script-evaluation organizer’s report
In another release discussion, one user values medium for a cost-sensitive workflow while others complain about minimal completion. We read original comments instead of treating the pinned GPT-generated summary as a sample statistic. Public reports do not independently authenticate accounts, routing or logs. We claim no community consensus and do not convert relative timestamps into unverified publication dates.
A writing-test organizer separately reports ten script tasks, five drafts each and four LLM judges. More reasoning did not improve that writing score, and judges disagreed. The exercise measures a particular channel’s voice and script requirements, not a general writing regression. Fluent prose, imitation of an author and factual accuracy need separate acceptance criteria.
Set effort for the task rather than always maximizing it
Terra’s API documentation lists none, low, medium by default, high, xhigh and max. Work/Codex UI settings and availability require their own documentation. Do not treat API parameter names as an exact mapping to Chat’s slider, or confuse Ultra’s multi-agent execution with single-model max.
ChatGPT Work and Codex models guide · OpenAI model-selection guide · GPT-6.1 Terra API model documentation · Reddit: script-evaluation organizer’s report
Start with the current product default and explicit acceptance criteria on a fixed task. If failure involves missed constraints, geometry or multistep planning, compare higher effort. If tools are disconnected, files inaccessible or permission denied, a larger reasoning budget will not repair the environment. For writing, blind-review tone and information density rather than rewarding length.
| Symptom | First action | Then compare |
|---|---|---|
| Missed reasoning constraints | Specify inputs and acceptance; retain failures | Whether more effort fixes the same model’s result |
| Unavailable tool/file | Check permission, connection and environment | Do not attribute an unresolved environment to intelligence |
| Verbose or wrong voice | Blind-review audience fit, information and style | Default versus lower effort |
| Early stop or false completion | Check artifacts, unfinished work and stop reason | Completion/rework cost in the same harness |
ChatGPT Work and Codex models guide · OpenAI model-selection guide · Reddit: unfinished-work early-stop report · Reddit: script-evaluation organizer’s report
Practice one: checkable data organization in Chat
Paste this synthetic CSV into a Chat model you can actually access, followed by the prompt. It requires no web action or account data. Duplicate IDs, a pending item and a missing amount are deliberate. The goal is not to sum every row but to preserve uncertainty and produce checkable JSON. To test image understanding, render the same table as an image yourself and retain the CSV as the reference.
After deduplication, the known paid amount should be USD 12.50 and the pending amount USD 7.25. A-003 remains unknown, not zero. Compare available default and higher thinking levels in fresh conversations with identical input and rules. We did not run these model tests; the output requirements define an acceptance contract.
invoice_id,amount,currency,status
A-001,12.50,USD,paid
A-002,7.25,USD,pending
A-001,12.50,USD,paid
A-003,,USD,paidProcess only the synthetic CSV below. Do not browse or make payments.
Deduplicate by invoice_id; report duplicate IDs without double-counting.
Separate paid from pending. Preserve missing amount as unknown, never0.
Output JSON only, with exactly these keys: currency,paid_total,pending_total,
duplicate_ids,missing_amount_ids,needs_review.
Use two-decimal strings for amounts, sorted ID arrays, and a boolean needs_review.
Check rows and statuses before calculating. Add no text outside the JSON.Practice two: verify deliverables in Work or Codex
For saved files and multistep execution, move to Work or Codex with suitable file/tool access and use the next prompt. All reads, writes and three deliverables remain inside one dedicated temporary experiment directory. If a tool is unavailable, report nonexecution instead of inventing paths or presenting a JSON reply as a saved file.
ChatGPT Work and Codex models guide · OpenAI model-selection guide
This separates completion from confident language: result JSON must satisfy the gate, an audit must explain deduplication and unknowns, and status must distinguish completion, blockers and untested work. Explain a blocked step specifically while continuing other authorized, safe, independent steps. That does not authorize bypassing approvals, identity checks or safety restrictions.
Process my fixture.csv inside the dedicated ./model-eval-lab directory.
Read and write only inside ./model-eval-lab. No network, accounts or other directories.
Apply the same deduplication, status and missing-value rules as the previous prompt.
Save:
./model-eval-lab/result.json
./model-eval-lab/audit.md
./model-eval-lab/status.json
Explain source evidence, deduplication and unknowns in the audit.
Distinguish completed/blocked/not_tested in status.
Use the saved offline gate if available; otherwise say validation was not executed.
If blocked, continue only other authorized safe independent steps; bypass no approval.
Report each genuinely saved path, verification result and unfinished requirement.Catch polished but wrong answers with an offline gate
This Python program checks fields, types and expected values for these four synthetic rows only. It is not a general financial validator. With no argument it tests one correct output and three errors: counting a duplicate twice, hiding a missing amount and merely claiming completion. We actually executed these four cases offline.
If you later save a real model’s result.json, supply its path as the script’s single argument. Acceptance returns 0 and a contract mismatch 2; malformed JSON raises a parse error. A process not crashing is not business correctness. The gate cannot detect requirements absent from its rules, so inspect the input, audit explanation and execution boundaries as well.
import copy
import json
import sys
from pathlib import Path
# Synthetic fixture only. This is not an accounting or general invoice validator.
EXPECTED = {
"currency": "USD",
"paid_total": "12.50",
"pending_total": "7.25",
"duplicate_ids": ["A-001"],
"missing_amount_ids": ["A-003"],
"needs_review": True,
}
def accepted(value):
if not isinstance(value, dict) or set(value) != set(EXPECTED):
return False
return all(type(value[k]) is type(v) and value[k] == v
for k, v in EXPECTED.items())
if len(sys.argv) == 2:
value = json.loads(Path(sys.argv[1]).read_text(encoding="utf-8"))
ok = accepted(value)
print(json.dumps({"accepted": ok}))
raise SystemExit(0 if ok else 2)
cases = [("expected", copy.deepcopy(EXPECTED), True)]
bad_total = copy.deepcopy(EXPECTED)
bad_total["paid_total"] = "25.00"
cases.append(("duplicate_counted_twice", bad_total, False))
bad_missing = copy.deepcopy(EXPECTED)
bad_missing["missing_amount_ids"] = []
cases.append(("missing_amount_hidden", bad_missing, False))
cases.append(("unfinished_summary", {"status": "done"}, False))
for name, value, wanted in cases:
assert accepted(value) is wanted, name
print(json.dumps({"offline_self_tests": len(cases), "passed": True}))Design paired comparisons instead of choosing a good screenshot
First pair the same model’s default and higher effort, then pair Terra and Sol. Avoid changing model, prompt, tools and budget together. Prepare several independent tasks per category, freeze acceptance criteria beforehand, rotate order and retain every outcome. Record completion, errors, timeouts, clarification and recovery instead of keeping only the best screenshot.
Cost includes actual consumption, retries, review and rework. Total cost per accepted deliverable equals all attempt costs divided by accepted count; if none pass, do not report a finite unit cost. Subscription-use percentages are not a simple conversion from API token prices. Small samples can guide the next experiment without becoming a precise market-wide ranking.
| Layer | Freeze or record | Criterion |
|---|---|---|
| Surface/configuration | Date, actual model, effort, tools, versions | Conditions reconstructable; unknowns not guessed |
| Task/artifacts | Input hash, allowed paths, JSON/file contract | Same inputs and gate |
| Run outcome | Pass, error, timeout, blocker, retry | Adverse outcomes retained |
| Cost/human work | Consumption, waiting, review, rework | Total cost per accepted deliverable |
Recent product changes introduce separate permissions and boundaries
Record product changes separately. Wallet began gradually rolling out in dots on November 13 to eligible adult Pro subscribers in the US, with eligible US-issued cards and billing addresses and review and authorization for each purchase. It is a controlled payment surface, not a new Terra capability score. Authorization does not establish merchant order confirmation. We analyzed documentation only and initiated no purchase.
ChatGPT product release notes · OpenAI ChatGPT Wallet announcement
The November 10 announcement schedules Advanced and Standard Voice retirement for December 11; GPT-Live is an existing voice direction. It retains an exception for live video and screen sharing to remain on Advanced voice until GPT-Live supports them. Images 2.6 on November 4 is an image-generation update. Different dates, modalities and eligibility must not become “every ChatGPT switched to a new model today.”
From product news to an accepted deliverable
- Record Chat/Work/Codex surface and actual model
- Separate vendor tests, independent tests and community signals
- Freeze input, permissions, budgets and acceptance contract
- Pair default/higher effort and candidate models
- Inspect artifacts, failures and total review cost
- Retain uncertainty and choose by task
Apply the evidence to everyday model choices
For conversation, rewriting and lightweight analysis, first verify facts, format and expression in an actually available Chat model, then decide whether more thinking is needed. For files, application control and successive deliverables, compare Terra’s default with Sol in Work/Codex while holding the tool environment fixed.
ChatGPT Work and Codex models guide · OpenAI model-selection guide · Artificial Analysis original Terra results · Bito independent coding tests
Terra’s value is not winning every leaderboard. It is completing a class of work sooner at an acceptable verification cost. Keep stronger-model review or human acceptance for complex, open-ended or consequential results. Reuse inputs, gates and logs after the next release to distinguish model progress from environment changes and better prompts.
This November16,2026 analysis combines public evidence with synthetic gate self-tests. We measured no personal account, subscription consumption, live API or community user’s internal run. Read model capabilities, product surfaces and task acceptance together to turn releases into verifiable choices.