GPT-5.6 Sol: Capabilities, ChatGPT Access, API Pricing, and What Changes From GPT-5.5
- 4 minutes ago
- 13 min read
OpenAI released GPT-5.6 Sol as the flagship model in a new three-tier family whose structure separates maximum professional capability from lower-cost production workloads, while GPT-5.5 Instant continues to serve as the default model for fast, everyday responses inside ChatGPT.
Although GPT-5.6 Sol preserves GPT-5.5’s standard API prices, one-million-token context window, and 128,000-token maximum output, it changes how the model reasons across difficult tasks, coordinates tools, preserves reasoning between requests, operates parallel agents, and produces finished documents, spreadsheets, presentations, interfaces, and software artifacts.
The generation change therefore concentrates on execution efficiency and workflow depth rather than on a larger numerical context limit, because OpenAI has designed GPT-5.6 to reach frontier performance with fewer output tokens while reducing the number of model turns and tool calls required by complex agentic work.
For ChatGPT users, access depends on the subscription plan and selected reasoning level, whereas API developers can choose among Sol, Terra, and Luna, each of which belongs to the same GPT-5.6 generation while occupying a different position in the capability, latency, and pricing range.
·····
GPT-5.6 Sol Leads a Three-Tier Model Family Whose Names Represent Persistent Capability Levels.
OpenAI has divided GPT-5.6 into Sol, Terra, and Luna, with Sol serving as the flagship model for complex professional work, Terra balancing capability and cost for everyday production workloads, and Luna emphasizing speed and lower operating expense for repetitive or high-volume processing.
The number identifies the model generation, while the names describe capability tiers that OpenAI intends to advance on their own cadence, which allows the company to improve a particular tier without abandoning the broader family structure or forcing customers to return to an older numbered generation.
Developers who request the gpt-5.6 alias receive GPT-5.6 Sol, whereas applications that require Terra or Luna must select those models explicitly, which makes Sol the default API destination for teams that replace the GPT-5.5 model identifier without redesigning their routing logic.
Because Terra is priced at half of Sol’s standard token rates while remaining competitive with GPT-5.5 on many workloads, organizations can move routine stages to Terra while reserving Sol for planning, synthesis, verification, difficult coding, and other operations whose failure cost exceeds the additional model expense.
Luna extends that routing model toward high-volume work, since its standard input and output rates are one fifth of Sol’s prices, although its lower capability makes it more appropriate for bounded transformations, classification, extraction, and other tasks whose instructions and acceptance criteria are already well defined.
........
GPT-5.6 Family Positioning and Standard Short-Context API Pricing.
Model | Position Within the Family | Appropriate Workloads | Input per 1M Tokens | Cached Input per 1M Tokens | Output per 1M Tokens |
GPT-5.6 Sol | Flagship frontier tier | Complex reasoning, coding, research, computer use, science, cybersecurity, design, and professional artifact production | $5.00 | $0.50 | $30.00 |
GPT-5.6 Terra | Balanced production tier | Everyday professional workflows that require substantial capability at a lower operating cost | $2.50 | $0.25 | $15.00 |
GPT-5.6 Luna | Fast and economical tier | Repetitive, latency-sensitive, cost-sensitive, and high-volume processing | $1.00 | $0.10 | $6.00 |
·····
The Context Window and Base Prices Remain Unchanged, Although the Knowledge Cutoff and Execution Controls Advance.
GPT-5.6 Sol provides a 1,050,000-token context window and supports as many as 128,000 output tokens, which matches GPT-5.5 and confirms that the new generation does not expand the maximum quantity of text and images that developers can place inside one API request.
Its documented knowledge cutoff advances to February 16, 2026, compared with December 1, 2025, for GPT-5.5, while both models support text and image input with text output and provide reasoning-token functionality for tasks that require internal analysis before the final response is produced.
Although the context ceiling remains stable, GPT-5.6 changes how applications use that capacity, because persisted reasoning can carry relevant reasoning items between calls, explicit prompt-cache breakpoints can isolate reusable prompt prefixes, and Programmatic Tool Calling can filter intermediate data before it occupies additional model context.
Long-context pricing also remains aligned with GPT-5.5, since requests containing more than 272,000 input tokens are charged at twice the normal input rate and one and a half times the normal output rate for the entire request rather than only for the tokens above the threshold.
The unchanged price card does not guarantee an unchanged invoice, because GPT-5.6 introduces separately billed cache writes while Pro execution, maximum reasoning, multi-agent coordination, external tools, and longer workflows can alter the number of tokens and metered operations consumed before an accepted result is produced.
........
Core Technical Differences Between GPT-5.5 and GPT-5.6 Sol.
Specification or Capability | GPT-5.5 | GPT-5.6 Sol |
Context window | 1,050,000 tokens | 1,050,000 tokens |
Maximum output | 128,000 tokens | 128,000 tokens |
Knowledge cutoff | December 1, 2025 | February 16, 2026 |
Standard API input price | $5.00 per 1M tokens | $5.00 per 1M tokens |
Standard API output price | $30.00 per 1M tokens | $30.00 per 1M tokens |
Highest reasoning effort | xhigh | max |
Premium API execution | Separate GPT-5.5 Pro model | Pro mode on the selected GPT-5.6 model |
Prompt caching | Cached-input discount without GPT-5.6 cache-write pricing | Explicit or implicit caching with separately billed cache writes |
Reasoning continuity | No corresponding GPT-5.6 control | Persisted reasoning through reasoning.context |
Programmatic Tool Calling | Not supported | Supported through the Responses API |
Native multi-agent beta | Not available | Concurrent subagents through the Responses API beta |
Flagship alias | gpt-5.5 | gpt-5.6, which routes to gpt-5.6-sol |
·····
Reasoning Levels, Pro Mode, and Multi-Agent Execution Change How Computation Is Assigned to Difficult Tasks.
GPT-5.6 supports none, low, medium, high, xhigh, and max reasoning efforts, with max extending beyond GPT-5.5’s previous upper level so that demanding requests can receive additional exploration, checking, revision, and verification when quality carries greater weight than latency or token consumption.
OpenAI recommends that developers migrating from GPT-5.5 retain their current reasoning setting as the first baseline and then test the same setting alongside one level lower, because GPT-5.6 may preserve or improve task quality while producing fewer output tokens.
A workflow that currently runs GPT-5.5 at high should therefore compare GPT-5.6 at both high and medium, whereas an application already using xhigh should test max only when the additional reasoning produces a measurable gain in accepted completions, factual accuracy, or reduced human correction.
The treatment of Pro execution also changes, because GPT-5.5 Pro is a separately priced model whose standard rates reach $30 per million input tokens and $180 per million output tokens, while GPT-5.6 keeps Sol, Terra, or Luna as the selected model and enables additional computation through reasoning.mode: "pro".
Since reasoning effort remains independent from Pro mode, developers can select a model tier and effort level according to the workload, although the additional model work performed in Pro mode generally increases latency and total token consumption even though the underlying model’s published token rates remain unchanged.
GPT-5.6 also introduces persisted reasoning, which allows available reasoning items to be reused across turns when reasoning.context is configured, thereby reducing the need to reconstruct every intermediate assumption and decision each time a long-running workflow sends another request.
Multi-agent execution extends the architecture beyond one sequential reasoning process, because the Responses API beta can run concurrent subagents and synthesize their work within a single request, while OpenAI’s Ultra configuration coordinates four agents by default and trades higher aggregate token consumption for stronger results and shorter time to completion on parallelizable tasks.
........
GPT-5.6 Reasoning and Execution Options.
Option | Execution Pattern | Appropriate Conditions | Principal Operational Consequence |
none or low | Minimal or limited internal reasoning | Latency-sensitive and well-bounded requests | Lower reasoning cost and faster responses |
medium | Balanced reasoning allocation | General professional work with moderate complexity | Balanced quality, latency, and token use |
high or xhigh | Extended internal exploration and verification | Complex coding, research, planning, and analytical work | Higher token use and longer execution |
max | Highest standard GPT-5.6 reasoning effort | Quality-first tasks that justify extensive checking and revision | Highest standard reasoning expenditure |
Pro mode | Additional model work before one final answer | Difficult requests for which reliability outweighs speed | Greater latency and token consumption without a separate model slug |
Multi-agent or Ultra | Several agents work concurrently before synthesis | Work that divides into independent research, coding, or analysis streams | Higher aggregate token use with potentially shorter elapsed time |
·····
Programmatic Tool Calling Changes the Economics of Long and Tool-Heavy Workflows.
Traditional agent workflows often return every search response, database record, file extract, shell output, or application state to the model before the next action begins, which can enlarge the context and create a separate model round trip for each deterministic operation.
Programmatic Tool Calling allows GPT-5.6 to write and run lightweight programs that coordinate eligible tools, process intermediate results, monitor progress, and determine subsequent operations as work proceeds, while retaining only the information that the model needs for its final reasoning.
When a workflow must filter thousands of records, combine several search results, inspect multiple files, or transform structured data, the hosted program can discard irrelevant material before it re-enters the model context, which reduces prompt growth and avoids spending model tokens on intermediate information that has no effect on the final answer.
The mechanism also reduces model turns when several bounded operations can be executed inside one program, although workflows in which each result changes the interpretation, authorization boundary, safety decision, or objective still require fresh model judgment between tool calls.
OpenAI states that Programmatic Tool Calling is compatible with Zero Data Retention configurations, while customer evaluations cited at launch reported fewer output tokens, fewer model turns, and shorter completion times across financial research, document analysis, and structured scene-building workflows.
The economic effect depends on the composition of the workflow, because reducing model tokens and round trips may lower the cost of orchestration even when separately priced tools, hosted environments, search calls, or other metered operations remain part of the completed request.
........
Conventional and Programmatic Tool Execution Compared.
Workflow Characteristic | Conventional Tool Calling | Programmatic Tool Calling |
Tool sequencing | The model selects and evaluates operations individually | Generated programs coordinate eligible tools |
Intermediate results | Results commonly return to the model after each operation | Results can be processed before they re-enter model context |
Context growth | Large outputs may repeatedly consume prompt capacity | Irrelevant intermediate data can be filtered in the runtime |
Model turns | Several stages may require separate model round trips | Multiple bounded operations can occur within one program |
Best-suited tasks | Ambiguous workflows requiring repeated model judgment | Deterministic, bounded, and tool-heavy workflows |
Zero Data Retention | Depends on the selected tools and configuration | Supported for Programmatic Tool Calling |
Cost structure | Model tokens, tool fees, and repeated orchestration | Model tokens and tool fees with fewer potential model turns |
·····
Coding, Computer Use, Science, and Cybersecurity Show the Largest Published Performance Gains.
OpenAI positions GPT-5.6 Sol as a model for extended engineering work in which planning, implementation, terminal use, debugging, testing, visual inspection, and revision must remain coordinated across many steps rather than being completed through one isolated code response.
On OpenAI’s published evaluations, GPT-5.6 Sol scores 64.6 percent on SWE-Bench Pro compared with 59.4 percent for GPT-5.5, while Terminal-Bench 2.1 rises from 85.6 percent to 88.8 percent for one Sol agent and reaches 91.9 percent under the Ultra multi-agent configuration.
Computer-use gains are larger on several evaluations, because OSWorld 2.0 rises from 47.5 percent with GPT-5.5 to 62.6 percent with Sol, while BenchCAD increases from 44.4 percent to 70.6 percent and BrowseComp moves from 84.4 percent to 90.4 percent for one Sol agent.
OpenAI also reports a 92.2 percent BrowseComp score for Sol Ultra, although that result uses parallel agents rather than the single-agent Sol configuration, which should be considered when organizations compare the benchmark improvement against the additional token consumption required by multi-agent execution.
Scientific evaluations show substantial improvements across biology, life-science research, and medicinal chemistry, with GeneBench Pro increasing from 12 percent to 28.7 percent, LifeSciBench moving from 50.4 percent to 59.9 percent, and OpenAI’s internal MedChemBench rising from 35.5 percent to 48.3 percent.
Cybersecurity results increase sharply under OpenAI’s evaluation conditions, since SEC-Bench Pro rises from 45.8 percent to 71.2 percent and ExploitBench reaches 73.5 percent compared with GPT-5.5’s 47.9 percent, while the release also expands support for secure code review, patching, threat modeling, blue teaming, vulnerability triage, and authorized defensive research.
Although the launch evaluations show broad gains, the improvement is not uniform across every test, because GPT-5.6 Sol scores 73.8 percent on OpenAI MRCR v2 between 512,000 and one million tokens compared with GPT-5.5’s 74 percent, while its one-million-token GraphWalks score increases from 45.4 percent to 77.1 percent.
........
Selected OpenAI Evaluations Comparing GPT-5.6 Sol With GPT-5.5.
Evaluation | GPT-5.5 | GPT-5.6 Sol | GPT-5.6 Sol Ultra |
SWE-Bench Pro | 59.4% | 64.6% | — |
Terminal-Bench 2.1 | 85.6% | 88.8% | 91.9% |
GeneBench Pro | 12.0% | 28.7% | — |
LifeSciBench | 50.4% | 59.9% | — |
OSWorld 2.0 | 47.5% | 62.6% | — |
BrowseComp | 84.4% | 90.4% | 92.2% |
BenchCAD | 44.4% | 70.6% | — |
SEC-Bench Pro | 45.8% | 71.2% | 74.3% |
ExploitBench | 47.9% | 73.5% | — |
OpenAI MRCR v2, 512K–1M | 74.0% | 73.8% | — |
GraphWalks BFS, 1M | 45.4% | 77.1% | — |
·····
Documents, Spreadsheets, Presentations, and Interfaces Receive Greater Structural and Visual Control.
GPT-5.6 extends its professional capabilities into artifact production, where layout, typography, spacing, hierarchy, formulas, charts, editable structure, and template fidelity determine whether an output can enter an operational workflow without extensive manual reconstruction.
When Sol receives a reference presentation, it can infer the deck’s design system from layouts, typography, spacing, colors, recurring content patterns, and rules embedded in the Slide Master, after which it can apply those conventions to newly generated slides with greater consistency than GPT-5.5.
Document and spreadsheet production follows the same direction, since GPT-5.6 is designed to preserve complex reference formats while handling equations, financial models, page composition, worksheet organization, visual hierarchy, and spacing with greater precision.
Frontend development combines code generation with rendered inspection, which allows the model to produce an interface, examine the live result, identify visual or functional defects, revise the implementation, and evaluate the updated version during the same extended workflow.
Customer evaluations cited by OpenAI reported that GPT-5.6 scored 4.4 on a five-point frontend quality rubric compared with 4.0 for GPT-5.5, while other evaluations reported fewer steps, fewer tool calls, reduced token consumption, and less rework across application development, presentation production, and design-to-code workflows.
These improvements change the point at which the model participates in professional production, because GPT-5.5 frequently supplied content or code that required downstream assembly, whereas GPT-5.6 is designed to inspect and refine the rendered artifact before handing it back to the user.
........
Professional Production Capabilities in GPT-5.6 Sol.
Workload | GPT-5.6 Sol Capability | Operational Consequence |
Presentations | Infers layouts, typography, colors, spacing, recurring structures, and Slide Master rules | Newly generated slides can follow an existing deck with fewer missing structural components |
Documents | Applies hierarchy, equations, typography, spacing, and page-layout conventions | Reports and professional documents require less manual reconstruction |
Spreadsheets | Handles formulas, financial models, worksheet layouts, and reference formatting | Generated workbooks remain editable while preserving analytical relationships |
Frontend development | Generates code, inspects rendered interfaces, and revises defects | Visual correction becomes part of the implementation workflow |
Computer use | Interacts with software interfaces while monitoring visible state | The model can validate work inside the application where the artifact is used |
Template-based production | Extracts recurring visual and structural rules from references | Repeated outputs can follow established organizational formats more consistently |
·····
ChatGPT Access Depends on the Plan, the Reasoning Level, and the Product Surface.
GPT-5.5 Instant remains the default model for fast responses in standard ChatGPT conversations, whereas GPT-5.6 Sol powers Medium, High, and Extra High reasoning on eligible plans and GPT-5.6 Sol Pro powers the Pro option for difficult tasks and longer-running workflows.
When automatic switching is enabled on an eligible paid account, ChatGPT can move from Instant to Medium when a request requires additional reasoning, while users can also choose an available reasoning level directly through the model picker.
Plus subscribers receive Medium and High access, while Pro, Business, and Enterprise accounts receive Medium, High, Extra High, and Pro access, although administrators in managed Business and Enterprise workspaces can restrict which models members are allowed to use.
Free, Go, and logged-out users do not receive GPT-5.6 Sol in standard ChatGPT conversations, although Free and Go users can access GPT-5.6 Terra through limited ChatGPT Work and Codex experiences where those products are available.
Terra and Luna are not selectable in ordinary ChatGPT conversations, whereas ChatGPT Work provides Sol, Terra, and Luna to Plus, Pro, Business, and Enterprise users, and Codex provides Terra to Free and Go users while offering all three tiers to eligible paid plans.
GPT-5.6 access is included according to the user’s existing ChatGPT plan rather than sold as a separate model subscription, while usage limits, regional billing, taxes, managed-workspace controls, and gradual rollout conditions determine how much access an individual account receives in practice.
........
GPT-5.6 Availability in Standard ChatGPT Conversations.
ChatGPT Plan | Medium and High With GPT-5.6 Sol | Extra High With GPT-5.6 Sol | GPT-5.6 Sol Pro |
Free | Not included | Not included | Not included |
Go | Not included | Not included | Not included |
Plus | Included | Not included | Not included |
Pro | Included | Included | Included |
Business | Included | Included | Included |
Enterprise | Included | Included | Included |
·····
API Pricing Preserves GPT-5.5’s Base Rates While Adding Explicit Cache-Write Billing.
GPT-5.6 Sol costs $5 per million standard input tokens, $0.50 per million cached input tokens, and $30 per million output tokens, which exactly matches GPT-5.5’s published standard short-context rates.
Requests containing more than 272,000 input tokens use long-context prices for the entire request, under which Sol costs $10 per million input tokens, $1 per million cached input tokens, and $45 per million output tokens.
GPT-5.6 introduces cache-write billing at 1.25 times the ordinary uncached input rate, while cached reads continue to receive a 90 percent discount, which makes the economics of caching dependent on how often the stored prefix is reused after its initial write.
Sol cache writes cost $6.25 per million tokens for short-context requests and $12.50 per million tokens for long-context requests, while explicit cache breakpoints allow developers to identify which stable prompt prefixes should be stored rather than relying exclusively on automatic cache placement.
Batch and Flex processing reduce the standard token prices by half, whereas Priority processing increases short-context Sol pricing to $10 per million input tokens, $1 per million cached input tokens, $12.50 per million cache-write tokens, and $60 per million output tokens.
Since GPT-5.5 Priority pricing reaches $12.50 per million input tokens and $75 per million output tokens, GPT-5.6 Sol provides lower Priority rates even though its normal standard rates remain unchanged from the preceding flagship model.
........
Standard API Pricing per One Million Tokens.
Model | Short Input | Short Cached Input | Short Cache Write | Short Output | Long Input | Long Cached Input | Long Cache Write | Long Output |
GPT-5.6 Sol | $5.00 | $0.50 | $6.25 | $30.00 | $10.00 | $1.00 | $12.50 | $45.00 |
GPT-5.6 Terra | $2.50 | $0.25 | $3.125 | $15.00 | $5.00 | $0.50 | $6.25 | $22.50 |
GPT-5.6 Luna | $1.00 | $0.10 | $1.25 | $6.00 | $2.00 | $0.20 | $2.50 | $9.00 |
GPT-5.5 | $5.00 | $0.50 | — | $30.00 | $10.00 | $1.00 | — | $45.00 |
GPT-5.5 Pro | $30.00 | — | — | $180.00 | $60.00 | — | — | $270.00 |
·····
Migration From GPT-5.5 Requires Workflow Measurement Rather Than a Direct Model Substitution.
The simplest API migration replaces gpt-5.5 with gpt-5.6-sol or the gpt-5.6 alias, although an unchanged prompt may still produce different behavior because GPT-5.6 tends to use fewer output tokens, coordinates tools differently, supports persisted reasoning, and provides additional reasoning and execution modes.
Prompts that instructed GPT-5.5 to be extremely brief may become too compressed under GPT-5.6, while prompts that manually described every tool step may prevent Programmatic Tool Calling from reducing orchestration, which requires migration testing to include prompt structure alongside the model identifier.
Applications that previously selected GPT-5.5 Pro must change their request structure because GPT-5.6 does not require a separate Pro API slug, which means that developers retain Sol, Terra, or Luna as the chosen model and activate additional computation through reasoning.mode: "pro".
Reasoning effort should be calibrated against representative production tasks, because raising every request to max may increase latency and token use without improving accepted output, while reducing an existing GPT-5.5 setting by one level may preserve quality when GPT-5.6 completes the same work more efficiently.
Caching also requires new cost monitoring, since explicit writes introduce a charge that becomes economical only when the cached prefix receives enough discounted reads, while rapidly changing system instructions or reference material may generate write costs without sufficient reuse.
Programmatic Tool Calling should be evaluated where several deterministic operations pass large intermediate outputs between tools, whereas multi-agent execution should be reserved for workloads whose components can progress independently and whose shorter elapsed time or higher completion rate offsets the additional aggregate token consumption.
The appropriate migration metric is therefore the cost of an accepted completed result, which includes input tokens, output tokens, reasoning expenditure, cache writes, cache reads, tool charges, retries, latency, human corrections, and failed runs rather than the published price of one million tokens considered in isolation.
For organizations that already use GPT-5.5 successfully, GPT-5.6 Sol offers the clearest operational gains when a workflow requires long reasoning chains, sustained tool use, rendered-output inspection, professional artifact generation, or parallel research, whereas stable and narrowly defined workloads may achieve lower costs by moving to Terra or Luna instead of defaulting every request to the flagship tier.
·····
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
·····
