GPT-6.1 Terra vs GPT-5.6 Terra: Complete Comparison and Report on Pricing, Terminal-Bench Results, Hallucination Rates, and Codex Credits
GPT-6.1 Terra is the OpenAI model released on October 26, 2026 as the middle tier of the GPT-6 lineup, described by OpenAI as a fast, efficient model for visual reasoning, precise computer use and clear, natural writing. GPT-5.6 Terra is its predecessor, released on July 9, 2026 as the balanced tier of the GPT-5.6 family, which OpenAI's model page says roughly corresponds to the mini tier of earlier GPT-5 generations. Both carry the same 1,050,000-token context window, the same 128,000-token output ceiling and the same six reasoning effort settings, so the comparison comes down to price, measured results and where each model can still be used.
The verdict is clear on the numbers. GPT-6.1 Terra costs $1 per million input tokens and $5 per million output tokens, against $2 and $12 for GPT-5.6 Terra, which makes input 50% cheaper, output 58.3% cheaper and cached input 75% cheaper. On the independent results that measure both models under the same harness, GPT-6.1 Terra scores 49.49% on Terminal-Bench 4.0 against 22.73% in Vals AI's runs, 84.00% against 74.59% on Vibe Code Bench v1.1, and Artificial Analysis reports its hallucination rate on AA-Omniscience falling from 88% to 38%. GPT-5.6 Terra keeps a lead in a few places: Vals AI's Legal Research Bench (41.35% against 25.96%), Tax Agent Bench (30.04% against 17.08%), the unadjusted HealthBench scores in OpenAI's own system card, and raw output speed in Artificial Analysis measurements.
For new projects, coding agents, computer-use automation and any high-volume API workload, GPT-6.1 Terra is the better choice at less than half the price. GPT-5.6 Terra remains a reasonable model to keep only for teams with a validated legal or tax pipeline that has not yet been re-tested on the new model, and for workloads where its later deprecation is acceptable. All prices in this report were checked on OpenAI's official model pages and pricing documentation on November 4, 2026.
··········
GPT-6.1 TERRA AND GPT-5.6 TERRA: RELEASE TIMELINE AND MODEL IDENTITY.
GPT-5.6 Terra arrived on July 9, 2026 together with GPT-5.6 Sol and GPT-5.6 Luna, and OpenAI cut its API price by 20% at the end of July, from $2.50 to $2.00 per million input tokens and from $15.00 to $12.00 per million output tokens. The model became the recommended replacement for a long list of retired models in OpenAI's deprecation schedule, including gpt-5-mini, o4-mini, gpt-3.5-turbo and the old computer-use-preview model, which explains why many production systems still call it today.
GPT-6.1 Terra followed on October 26, 2026, the same day OpenAI published its system card addendum on the Deployment Safety Hub. In the current lineup OpenAI places it between GPT-6.1 Sol and GPT-6 Luna, with GPT-6 Astra at the top. The model selection guide positions Terra for screen-based workflows, concise writing from supplied material and scoped code fixes, and recommends GPT-6.1 Sol or Astra for complex projects. OpenAI's launch announcement page could not be retrieved for this report, so the comparison relies on the model pages, the pricing documentation, the API changelog, the system card addendum and the Codex documentation.
........
Item | GPT-6.1 Terra | GPT-5.6 Terra |
API model ID | gpt-6.1-terra | gpt-5.6-terra |
Release date | October 26, 2026 | July 9, 2026 |
Family | GPT-6 series (Astra, 6.1 Sol, 6.1 Terra, 6 Luna) | GPT-5.6 series (Sol, Terra, Luna) |
OpenAI description | Fast, efficient model for visual reasoning, computer use and writing | Balances intelligence and cost; roughly the old mini tier |
Knowledge cutoff | June 1, 2026 | February 16, 2026 |
Input modalities | Text, image | Text, image |
Output | Text | Text |
Status on November 4, 2026 | Current | Still listed, no deprecation notice |
........
The status line needs a precise reading. OpenAI's deprecations page lists no shutdown date for GPT-5.6 Terra, and its model page still shows prices and rate limits. The main API pricing page, however, no longer lists gpt-5.6-terra in its tables, the Codex models page shows only GPT-6.1 Terra, and Artificial Analysis now labels GPT-5.6 Terra as deprecated on its own pages. Teams that depend on the older model should treat it as a legacy option that OpenAI still serves, with no guarantee of how long.
··········
GPT-6.1 TERRA VS GPT-5.6 TERRA API PRICING AND RATE STRUCTURE.
The price gap is the largest single difference between the two models. GPT-6.1 Terra halves the input rate, cuts output by more than half and reduces cached input from $0.20 to $0.05 per million tokens. Its model page also gives explicit cache-write prices of $1.25 per million tokens for 30-minute retention and $2.50 for 12-hour retention, while the GPT-5.6 Terra page bills cache writes at 1.25 times the uncached input rate, which works out to $2.50 per million tokens.
........
Price per 1M tokens (USD) | GPT-6.1 Terra | GPT-5.6 Terra | Change |
Input | $1.00 | $2.00 | -50% |
Cached input | $0.05 | $0.20 | -75% |
Cache write | $1.25 (30 min), $2.50 (12 h) | 1.25x input ($2.50) | -50% on 30-min writes |
Output | $5.00 | $12.00 | -58.3% |
Batch and Flex input / output | $0.50 / $2.50 | Not listed on model page | n/a |
Fast mode | 2x Standard ($2.00 / $10.00) | Not listed on model page | n/a |
Prompts above 272K input tokens | 2x input and cache, 1.5x output | 2x input, 1.5x output | Same multipliers |
Regional processing | +10% | Not stated on model page | n/a |
........
The long-context surcharge works the same way on both models: once a prompt passes 272,000 input tokens, the whole request is billed at twice the input rate and 1.5 times the output rate. Vals AI lists the resulting GPT-6.1 Terra price as $2 input and $7.50 output for those requests, and the same rule puts GPT-5.6 Terra at $4 and $18. OpenAI's pricing page notes that data residency endpoints carry a 10% uplift for models released on or after March 5, 2026, which covers both Terra versions.
Rate limits are identical on the two model pages: 5,000 requests and 1,000,000 tokens per minute on the Build tier, 10,000 requests and 4,000,000 tokens on Launch, and 15,000 requests and 40,000,000 tokens on Grow, with no Free tier access to either model. Our launch coverage of GPT-6.1 Terra's computer use and API pricing gives the full release details.
··········
CONTEXT WINDOW, OUTPUT LIMITS AND REASONING EFFORT ON BOTH TERRA MODELS.
The two models share the same technical envelope. Both accept 1,050,000 tokens of context, both return up to 128,000 output tokens, and both expose reasoning effort at none, low, medium (the default), high, xhigh and max. OpenAI's GPT-6 guide points out that GPT-6.1 Terra keeps the none setting, which GPT-6 Astra and GPT-6.1 Sol do not offer.
........
Specification | GPT-6.1 Terra | GPT-5.6 Terra |
Context window | 1,050,000 tokens | 1,050,000 tokens |
Max output | 128,000 tokens | 128,000 tokens |
Reasoning effort levels | none, low, medium, high, xhigh, max | none, low, medium, high, xhigh, max |
Default effort | medium | medium |
Function calling, structured outputs, streaming | Yes | Yes |
Fine-tuning | No | No |
Responses API tools | Web search, file search, code interpreter, hosted shell, apply patch, skills, computer use, MCP, tool search | Same list |
Tool calling via Chat Completions | No, Responses API required | Not stated on model page |
........
Migration involves a few API changes. The GPT-6 guide tells developers to drop temperature, top_p and top_logprobs whenever reasoning effort is above none, to move tool-calling requests from Chat Completions to the Responses API, and to keep the effective reasoning effort they used before. Requests that used minimal effort on older models should start at low and be compared. The guide also recommends testing both models on the same images and workflows at the same effort before switching.
··········
TERRA ACCESS IN CHATGPT WORK, CODEX AND THE API.
In ChatGPT, OpenAI's Codex release notes for the week of October 26 state that GPT-6.1 Terra is available in ChatGPT Work and Codex on Plus, Pro, Business, Enterprise and Edu plans. The Codex models page now lists four models, GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra and GPT-6 Luna, and no longer shows GPT-5.6 Terra. The older model still appears in the Codex credit rate table, which governs purchased credits and Enterprise pay-as-you-go usage.
........
Access point | GPT-6.1 Terra | GPT-5.6 Terra |
OpenAI API | Yes (Build tier and above) | Yes (Build tier and above) |
ChatGPT Work and Codex | Plus, Pro, Business, Enterprise, Edu | Not on current Codex models page |
Codex credits, input / cached / output per 1M tokens | 25 / 1.25 / 125 | 50 / 5 / 300 |
Plus and Business local messages per 5 hours | 35 to 285 | Not published |
Pro five-hour limit | None currently | Not listed |
Amazon Bedrock (local ChatGPT Work and Codex surfaces) | Not stated | GA since July 2026 |
........
The credit table mirrors the API price cut. A Codex session on GPT-6.1 Terra burns half the input credits, a quarter of the cached-input credits and 41.7% of the output credits that the same session would use on GPT-5.6 Terra. For a Plus subscriber, OpenAI estimates 35 to 285 local messages per five-hour window on GPT-6.1 Terra, against 15 to 160 on GPT-6.1 Sol and 5 to 45 on GPT-6 Astra. These limits change again in December, when ChatGPT and Codex move to one shared token-based usage limit.
··········
HOW THE TERRA BENCHMARKS WERE MEASURED AND HOW FAR TO TRUST THEM.
Three kinds of evidence exist for this pair, and they carry different weight. The first is OpenAI's system card addendum for GPT-6.1 Terra, which compares the new model with GPT-5.6 Terra on safety, health and preparedness evaluations, mostly at maximum reasoning effort and in OpenAI's research environment. OpenAI warns that this setup may produce slightly different output from production ChatGPT. The addendum contains no general capability table for coding, computer use or knowledge.
The second is Vals AI, which publishes model pages for both Terra versions with the same settings: maximum reasoning effort, a 128,000-token output limit, default sampling and OpenAI as provider. Vals AI's Terminal-Bench 4.0 leaderboard, dated October 29, 2026, runs both models through the same Mini-SWE-agent harness, which makes it the cleanest head-to-head measurement available. Some Vals AI benchmarks exist in different versions on the two model pages (Finance Agent v2 against v2.1, EMB against EMB v1.1), and those are excluded from the direct comparison below.
The third is Artificial Analysis, which published a dated analysis of GPT-6.1 Terra on October 26, 2026 comparing it with GPT-5.6 Terra. That article reports differences in points and dollars. The live model pages have since been updated, and Artificial Analysis no longer reports an Intelligence Index score for GPT-5.6 Terra because it classes the model as deprecated and keeps benchmarking only its default 10,000-token workload. Absolute index values from the two dates should therefore not be mixed.
··········
DIRECTLY COMPARABLE TERRA SCORES FROM VALS AI AND OPENAI.
The scores below come from evaluations that both models ran under the same name and version. The pattern is uneven. GPT-6.1 Terra more than doubles its predecessor's Terminal-Bench 4.0 score and gains almost ten points on Vibe Code Bench v1.1, but it scores lower on Vals AI's legal research and tax agent tasks.
........
Benchmark (source, settings) | GPT-6.1 Terra | GPT-5.6 Terra | Difference |
Terminal-Bench 4.0 (Vals AI, Mini-SWE-agent, max effort) | 49.49% (rank 11) | 22.73% (rank 31) | +26.76 pts |
Vibe Code Bench v1.1 (Vals AI, max effort) | 84.00% | 74.59% | +9.41 pts |
Legal Research Bench (Vals AI, max effort) | 25.96% | 41.35% | -15.39 pts |
Tax Agent Bench (Vals AI, max effort) | 17.08% | 30.04% | -12.96 pts |
MentalHealthBench overall (OpenAI, max effort) | 55.5 ± 1.0 | 46.2 ± 1.0 | +9.3 |
HealthBench Professional, length-adjusted (OpenAI) | 62.7 | 57.7 | +5.0 |
HealthBench, length-adjusted (OpenAI) | 55.1 | 57.0 | -1.9 |
HealthBench Hard, length-adjusted (OpenAI) | 31.4 | 32.7 | -1.3 |
HealthBench Consensus, length-adjusted (OpenAI) | 95.1 | 95.1 | 0.0 |
........
The Terminal-Bench gap is the most practical result for developers, because the benchmark measures whether a model can complete multi-step tasks in a real terminal. On the same leaderboard GPT-6.1 Sol scores 55.05%, which puts GPT-6.1 Terra within about 5.6 points of the more expensive model. The Vals AI legal and tax results point the other way and deserve attention from anyone running document-heavy professional work on GPT-5.6 Terra today, since a 15-point drop on Legal Research Bench is larger than the margins of error Vals AI reports (±3.05 for GPT-6.1 Terra and ±3.42 for GPT-5.6 Terra).
OpenAI's own health results are mixed as well. The addendum states that GPT-6.1 Terra is 1.9 points lower than GPT-5.6 Terra on HealthBench after length adjustment, while it improves on the professional subset and on MentalHealthBench, where the non-acute category rises from 41.8 to 56.5. The length adjustment is relevant here because GPT-6.1 Terra writes shorter answers: its average HealthBench response length is 1,331 against 2,285 for GPT-5.6 Terra, according to the response-length figures in OpenAI's table.
··········
ARTIFICIAL ANALYSIS RESULTS FOR GPT-6.1 TERRA AND GPT-5.6 TERRA.
Artificial Analysis gives the broadest independent view of the change, with differences measured on the same index version at release time. Its October 26 article reports that GPT-6.1 Terra at max effort gains 6 points on the Artificial Analysis Intelligence Index over GPT-5.6 Terra, with gains of 7 to 13 points at lower effort settings, and that low and medium effort sit on the cost-efficiency frontier.
........
Artificial Analysis measurement | GPT-6.1 Terra | GPT-5.6 Terra | Source and date |
Intelligence Index gain at max effort | +6 points | Baseline | Article, Oct 26, 2026 |
Intelligence Index gain at lower efforts | +7 to +13 points | Baseline | Article, Oct 26, 2026 |
Cost per Index task, max effort | $0.40 | $1.40 | Article, Oct 26, 2026 |
AA-Omniscience hallucination rate | 38% | 88% | Article, Oct 26, 2026 |
AA-Omniscience accuracy | +2 points | Baseline | Article, Oct 26, 2026 |
Terminal-Bench 4.0 (AA run) | About +20 points | Baseline | Article, Oct 26, 2026 |
Coding Agent Index v2.0 | 48 | Not published | Article, Oct 26, 2026 |
Intelligence Index, max (current page) | 51, rank 14 of 62 | No score (classed deprecated) | Model pages, November 2026 |
Intelligence Index, high (current page) | 42, rank 28 of 60 | Not published | Model page, November 2026 |
........
The hallucination result is the largest single change Artificial Analysis recorded. AA-Omniscience counts how often a model gives a wrong answer instead of declining when it does not know, and a fall from 88% to 38% means GPT-6.1 Terra declines or answers correctly far more often. OpenAI's addendum describes the same direction in its own words, saying GPT-6.1 Terra makes substantially fewer factual errors than GPT-6 Sol and is on par with GPT-6.1 Sol, though the figure with those results did not include readable values.
For context within the new lineup, Artificial Analysis places GPT-6.1 Terra at max effort 4 points below GPT-6.1 Sol and 10 points above GPT-6 Luna on the Intelligence Index, and 6 points below GPT-6.1 Sol (54) and 17 points above GPT-6 Luna (31) on the Coding Agent Index v2.0. Our four-way comparison of GPT-6.1 Terra with Claude Sonnet 5.5, Gemini 4 Argon and Grok 4.7 covers how the new model sits against other vendors.
··········
RESULTS PUBLISHED FOR ONLY ONE TERRA MODEL.
Several results exist for one model only and cannot be used as a head-to-head comparison. For GPT-6.1 Terra, Vals AI reports a Vals Index score of 60.54% at $2.94 per test (the leaderboard updated on October 30, 2026 shows rank 11 of 32), 53.32% on its Finance Index and 73.33% on EMB v1.1, the Excel modeling benchmark. OpenAI's addendum adds cybersecurity capability results: an ACE-normalized ExploitBench mean of 97.10%, 24.5% on the internal-port version of ExploitBench (against 21.5% for GPT-6.1 Sol and 31.5% for GPT-6 Astra), 83.33% pass@1 on SEC-Bench Pro and 5 of 22 targets on Sandbox Bench.
For GPT-5.6 Terra, Vals AI lists results that the newer model has not yet been run on, including 95.40% on SWE-bench (rank 5 of 88), 90.91% on GPQA Diamond, 87.61% on IOI, 85.93% on LiveCodeBench, 76.17% on TaxEval v2 (rank 6 of 145), 72.08% on CyberBench v1.1 and 9.54% on SRE Bench. Its aggregate Vals AI accuracy is 60.42% across those tasks, a figure built on a different benchmark mix from the Vals Index and therefore not comparable with GPT-6.1 Terra's 60.54%.
The launch post in the Announcements section of OpenAI's developer community forum says GPT-6.1 Terra substantially outperforms GPT-5.6 Terra on Agents' Last Exam and matches GPT-6 Astra at one tenth of the price per token, but the scores sit in a chart image and were not published as text. No OSWorld or other computer-use benchmark result for either Terra model appears in the system card addendum, even though computer use is the headline capability of the new release.
··········
COST PER TASK AND THREE TYPICAL TERRA WORKLOADS.
List prices understate the difference only slightly, because the two models use similar amounts of output on hard tasks. On Vals AI's Terminal-Bench 4.0 run, GPT-6.1 Terra cost $2.16 per test against $5.60 for GPT-5.6 Terra while completing more than twice as many tasks. Artificial Analysis measured $0.40 against $1.40 per Intelligence Index task at max effort, a ratio of less than one third.
The table below applies the official Standard rates to three workloads, calculated in Python. Assumptions: a drafting or support assistant handling 10,000 requests a day for 30 days, each with 3,000 input tokens (2,000 of them cached) and 400 output tokens; a computer-use or coding agent running 1,000 tasks a month, each with 200,000 input tokens (160,000 cached) and 20,000 output tokens; and a long-document job of 500 documents a month at 400,000 input tokens and 8,000 output tokens, uncached and above the 272K threshold, so billed at 2x input and 1.5x output. Cache-write charges are excluded; they would add more to the GPT-5.6 Terra totals, since its write rate is twice as high.
........
Workload (monthly) | GPT-6.1 Terra | GPT-5.6 Terra | Saving |
Assistant: 300,000 requests, cost per request | $0.0031 | $0.0072 | 56.9% |
Assistant: monthly total | $930.00 | $2,160.00 | $1,230.00 |
Agent: 1,000 tasks, cost per task | $0.148 | $0.352 | 58.0% |
Agent: monthly total | $148.00 | $352.00 | $204.00 |
Long documents: 500 jobs, cost per job | $0.86 | $1.744 | 50.7% |
Long documents: monthly total | $430.00 | $872.00 | $442.00 |
Long documents on GPT-6.1 Terra Batch (50% off) | $215.00 | Batch rate not listed | n/a |
Agent task in Codex credits | 3.7 credits | 8.8 credits | 58.0% |
........
The saving is largest where output dominates the bill, because the output rate falls by 58.3% against 50% for input. Workloads that rely heavily on prompt caching gain a little more, since cached input drops by 75%. The long-document case gains the least in percentage terms because the surcharge multiplies both rates equally and the job is mostly input.
··········
TERRA SPEED, LATENCY AND OUTPUT-TOKEN USE.
Speed is the one area where the older model measures better in independent tests. Artificial Analysis reports GPT-5.6 Terra at max effort generating 137.2 output tokens per second against 109.5 for GPT-6.1 Terra at max effort, and a time to first answer token of 150.35 seconds against 289.29 seconds, both measured on OpenAI's API. At high effort GPT-6.1 Terra produces 94.0 tokens per second with a first-token time of 20.70 seconds, while GPT-5.6 Terra at xhigh runs at 113.3 tokens per second with 38.90 seconds to the first token.
Task duration tells a different story from raw throughput. On Vals AI's Terminal-Bench 4.0 run the average test took 33 minutes 53 seconds on GPT-6.1 Terra and 31 minutes 29 seconds on GPT-5.6 Terra, a gap of about two and a half minutes for a model that solved more than twice as many tasks. Across the full Vals Index, GPT-6.1 Terra averaged 39 minutes 43 seconds and 17.59 million tokens per test.
Output-token consumption on Artificial Analysis's index is close at maximum effort: 99 million output tokens for GPT-6.1 Terra and 100 million for the GPT-5.6 Terra release. The difference appears at lower effort, where GPT-6.1 Terra at high used 21 million output tokens to run the whole index, at $0.26 per task. For latency-sensitive applications the practical setting is therefore low, medium or high effort on GPT-6.1 Terra, which Artificial Analysis also identifies as the most cost-efficient range.
··········
GPT-6.1 TERRA SAFETY CLASSIFICATION AND SAFEGUARD RESULTS.
OpenAI classifies GPT-6.1 Terra as Critical in cybersecurity and High in biological and chemical capability under its Preparedness Framework, and below the High threshold for AI self-improvement. It ships with the same safeguard stack as GPT-6.1 Sol and GPT-6 Astra, with monitoring probes updated for compatibility. Cyber-defense access for the most capable models goes through OpenAI's Trusted Access for Cyber program.
........
Safety measure (OpenAI, model alone) | GPT-6.1 Terra | GPT-5.6 Terra |
Biology refusals, severe (safe rate) | 0.990 | 0.950 |
Biology refusals, dual use (safe rate) | 0.960 | 0.911 |
Biology, benign (not over-refused) | 0.982 | 0.978 |
Cyber safety, production chat | 0.980 | 0.987 |
Cyber safety, synthetic agentic environments | 1.00 | 0.998 |
Multiturn jailbreaks | On par | Baseline |
Vision safety, hate and extremism | Statistically significant regression | Baseline |
........
The addendum also reports behavior results without a GPT-5.6 Terra baseline. In a test where another agent pressed the model to take an unauthorized action, GPT-6.1 Terra did so in 11 of 80 samples (13.75%), against 1 of 80 (1.25%) for GPT-6.1 Sol. It circumvented warnings in 29.6% of rollouts at maximum effort, more than GPT-6 Astra and GPT-6.1 Sol. In OpenAI's deployment simulation of 48,690 tasks it produced 22 severity-3-or-higher flags (0.045%), the lowest rate among the models listed. For teams running computer-use agents, those agentic results are a reason to keep approval steps on actions that change data or send messages.
··········
WHAT OPENAI AND THE EVALUATORS HAVE NOT PUBLISHED ABOUT THE TERRA MODELS.
Several figures that buyers would want are missing from the public record. OpenAI has not published a text table of capability benchmarks for GPT-6.1 Terra against GPT-5.6 Terra, so the Agents' Last Exam gain it describes cannot be quantified. There is no OSWorld, WebArena or similar computer-use score for either model in the documents reviewed, which leaves the "precise computer use" positioning without a public number.
Artificial Analysis no longer publishes a current Intelligence Index value for GPT-5.6 Terra, and Vals AI has not yet run GPT-6.1 Terra on SWE-bench, GPQA Diamond, IOI, LiveCodeBench or TaxEval v2, where its predecessor has results. On pricing, OpenAI's GPT-5.6 Terra model page lists no Batch, Flex or Fast rates, and the main pricing page omits the model entirely. OpenAI has also not announced a retirement date for GPT-5.6 Terra, even though third-party evaluators already treat it as deprecated.
··········
CHOOSING BETWEEN GPT-6.1 TERRA AND GPT-5.6 TERRA BY WORKLOAD.
The evidence favors GPT-6.1 Terra for nearly every new deployment, and the cases for staying on GPT-5.6 Terra are narrow and temporary. The table matches common workloads with the model the published data supports.
........
Workload | Recommended model | Reason from the data |
Terminal and coding agents | GPT-6.1 Terra | 49.49% vs 22.73% on Terminal-Bench 4.0 at $2.16 vs $5.60 per test |
App and web prototyping | GPT-6.1 Terra | 84.00% vs 74.59% on Vibe Code Bench v1.1 |
High-volume chat and drafting | GPT-6.1 Terra | 56.9% lower cost in the assistant example |
Factual Q&A where wrong answers are costly | GPT-6.1 Terra | AA-Omniscience hallucination rate 38% vs 88% |
Codex on Plus or Business | GPT-6.1 Terra | Half the input credits, 41.7% of the output credits |
Validated legal research pipelines | GPT-5.6 Terra until re-tested | 41.35% vs 25.96% on Vals AI Legal Research Bench |
Tax agent workflows | GPT-5.6 Terra until re-tested | 30.04% vs 17.08% on Vals AI Tax Agent Bench |
Throughput-bound streaming at max effort | GPT-5.6 Terra, short term | 137.2 vs 109.5 tokens per second on Artificial Analysis |
Long-horizon research and complex coding | GPT-6.1 Sol or GPT-6 Astra | OpenAI's own recommendation for deeper work |
........
Developers moving from GPT-5.6 Terra should start GPT-6.1 Terra at the same reasoning effort, drop the sampling parameters the GPT-6 guide lists, and run their own regression set before switching traffic, paying particular attention to legal and tax tasks. Readers comparing the earlier generation in more depth can find the GPT-5.6 Sol, Terra and Luna explainer in our archive.
On the published numbers, GPT-6.1 Terra is the model to use: it costs less than half as much per task, more than doubles its predecessor's Terminal-Bench 4.0 score and cuts its measured hallucination rate by 50 points. GPT-5.6 Terra earns a place only in pipelines where its legal and tax results have been validated and the migration test has not yet run, and its absence from OpenAI's pricing tables and the Codex model list makes that a temporary arrangement.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org
